Methods for controlling virtual assistants used in industrial workshops

Abstract

The present invention relates to a method for controlling a virtual assistant (10) for an industrial workshop, comprising: receiving (S10) an information request (Ir) by an input interface (20), wherein the information request (Ir) includes at least one request for receiving information about at least a portion of the industrial workshop; determining (S20) a model specification (Ms) by a control unit (30) using the received information request (Ir); determining (S30) a machine learning model (M) by a model manager (40) using the model specification (Ms); and providing (S40) a response (R) to the information request (Ir) by the control unit (30) using the determined machine learning model (M).

Description

TECHNICAL FIELD

[0001] The present invention relates to a method for controlling a virtual assistant for an industrial plant and a virtual assistant. BACKGROUND

[0002] Typically, plant operations are highly automated activities today. However, various situations require human intervention, which includes maintenance / support procedures, interventions to improve efficiency / quality, or to solve anomalies in the process.

[0003] These situations have information requirements: the operator needs data, in particular predictions. Which data and predictions are needed is typically unforeseeable during design and initial configuration of the system. Mechanisms are needed that allow automatic generation of corresponding data analysis and predictions.

[0004] Operators have information needs that are unforeseeable during design of the control system.

[0005] These needs arise during operation and can include estimation of time series values, anomaly detection of different data types with various types of pattern matching. This type of data analysis can be considered advanced, as it particularly tends to require analytical expertise and sometimes effort to select models and produce corresponding selections. Activities are difficult in settings where topology information and asset identifiers need to be set up in relation to sensors and corresponding historical data of the plant.

[0006] A consequence of the complexity is that information needs typically do not match predictions based on matching of analytical models to the current situation, but are mainly resolved by experience and intuition applied to the lifestream of sensor data.

[0007] While virtual assistants for industrial applications are used to help operators, the above problems have not been solved. SUMMARY

[0008] It is therefore an object of the present invention to provide an improved method for controlling a virtual assistant for an industrial plant. This object is achieved by the method according to claim 1 and the virtual assistant according to claim 15.

[0009] Further preferred embodiments are evident from the dependent patent claims.

[0010] According to the present application, a method for controlling a virtual assistant of an industrial plant comprises receiving, by an input interface, an information request, wherein the information request comprises at least one request for receiving information about at least a part of the industrial plant. The method further comprises determining, by a control unit, a model specification using the received information request. The method further comprises determining, by a model manager, a machine learning model using the model specification. The method further comprises providing, by the control unit, a response to the information request using the determined machine learning model.

[0011] Preferably, the input interface comprises a natural language interface and / or a dynamic user interface. In other words, the virtual assistant receives an information request from a user in a natural language form of the user or an information request entered via a graphical user interface, GUI.

[0012] The response to the information request of the user preferably comprises a process variable value, an event and / or an alarm of at least a part of the industrial plant.

[0013] Thus, an improved method for controlling a virtual assistant is provided.

[0014] In a preferred embodiment, the method comprises identifying, by the control unit, an information intent using the received information request and determining the model specification using the information intent.

[0015] Thus, the machine learning model is determined based on the information intent.

[0016] As used herein, the term “information intent” relates to an intent of a user formulating a request to the virtual assistant. The information intent preferably comprises a request option. The request option is preferably a predetermined request option. Further, the information intent preferably comprises a formalized declaration of the intent. Further, the information intent preferably comprises a specified response expectation.

[0017] Preferably, the information intent indicates a plant component of the industrial plant to be addressed and an information demand related to the plant component. For example, the information intent covers information on upcoming behavior of a specific plant component, e.g. a tank, in a specific location, e.g. in a specific sector. For example, a user enters the information intent into the virtual assistant by saying “Tell me about upcoming behavior of the tank in sector AB123”. Other examples include: “Estimate when the tank in sector ABC123 reaches a fill level of 20%”; “Predict the time until the temperature in plant section B reaches 50°C”.

[0018] Further, the information intent preferably comprises a list of tasks that have to be processed. In particular, such a list of tasks can comprise complex preconditions that have to be checked.

[0019] Preferably, using information intent to determine model specifications includes decomposing information intent into model specifications.

[0020] Preferably, determining the model input includes using topological information of the industrial workshop and information intent to map workshop components to corresponding sensors within the industrial workshop. In other words, the topological information is used to identify the sensors that need to be read out in order to determine the response.

[0021] The term "model specification" used here refers to the functionality of a machine learning model. In other words, a model specification indicates the technical requirements for a machine learning model that can determine the response to the intended message.

[0022] Considering the user's intent allows them to operate better by increasing contextual awareness provided by the virtual assistant.

[0023] Preferably, data analysis and / or predictions are provided to the user in relation to the information request related to the user's intent to use the information.

[0024] Therefore, an improved method for controlling virtual assistants is provided.

[0025] In a preferred embodiment, determining the machine learning model includes: using model specifications to check whether a suitable machine learning model is stored in the model database.

[0026] In other words, based on the model specification, a request is generated, requesting a machine learning model from the model manager that meets the model specification, and can therefore be used to determine the response to the information request, especially taking into account the information intent.

[0027] The term “suitable machine learning model” as used herein refers to a machine learning model that can be used to determine a response to an information request, especially taking into account the information intent.

[0028] Therefore, an improved method for controlling virtual assistants is provided.

[0029] In a preferred embodiment, if a suitable machine learning model is determined to be stored in a model database, the method includes the step of determining a response to an information request by using the stored machine learning model.

[0030] Therefore, users' information requests can be responded to in real time.

[0031] Therefore, an improved method for controlling virtual assistants is provided.

[0032] In a preferred embodiment, determining a response to an information request includes providing model input by a control unit, wherein the model input is determined by using information intent, and the response is determined by feeding the model input into a machine learning model.

[0033] In a preferred embodiment, if it is determined that there is no suitable machine learning model to be stored in the model database, the method includes the step of providing a delayed response to the user.

[0034] Preferably, the delayed response includes an estimated delay time. This allows the user to be notified in real time how long they must wait for a response to their information request.

[0035] Therefore, an improved method for controlling virtual assistants is provided.

[0036] In a preferred embodiment, if it is determined that no suitable machine learning model is stored in the model database, the method includes the following steps: determining machine learning model candidates using informational intent via an autoML pipeline; testing the model quality of the machine learning model candidates; if the model quality is acceptable, using the machine learning model candidates to determine a machine learning model; and if the model quality is unacceptable, notifying the user that model generation was unsuccessful.

[0037] In other words, in the autoML pipeline, information intent is used to determine the parameterization of machine learning model candidates.

[0038] Preferably, the machine learning models in the model database are labeled, particularly with respect to their functionality. In other words, the model specifications are compared with the labels of the machine learning models in the model database.

[0039] In other words, the use of an autoML pipeline depends directly on the intent of the information provided. For example, the intent of the information includes information on the frequency and / or temporal relevance of a particular machine learning model being used; that is, how often a particular machine learning model is used in a particular time frame.

[0040] The automatic generation of machine learning models based on information intent allows for the automation of interactions that would otherwise require time-consuming manual interaction.

[0041] Therefore, an improved method for controlling virtual assistants is provided.

[0042] In a preferred embodiment, identifying the intent of the information includes converting the information request into a machine-understandable format.

[0043] In other words, decomposing the information intent involves converting the information request into a machine-understandable format. For example, if the information request is in the form of the user's natural language, then the user's natural language is converted into a machine-understandable format that the control unit can process.

[0044] Therefore, an improved method for controlling virtual assistants is provided.

[0045] In a preferred embodiment, providing a response includes converting the determined response into a user-understandable format.

[0046] Therefore, an improved method for controlling virtual assistants is provided.

[0047] In a preferred embodiment, the method includes the following steps: a session manager determines user action information related to the actions and context information of the tracked individual user; the session manager uses the received user action information to determine individual user information; a state manager requests service overview information based on the individual user information; the state manager uses the received service overview information to determine global information; and the session manager uses the global information to determine the user's response.

[0048] Preferably, the determined individual user information is provided to the state manager.

[0049] Preferably, the determined global information is provided to the session manager.

[0050] Preferably, the session manager tracks individual users' actions and contextual information to determine user action information.

[0051] Preferably, the user action information includes the user's information request.

[0052] User action information is preferably received via an input interface.

[0053] The term "global information" as used herein includes information about ongoing processes, workflows, tasks, and actions on the industrial shop floor. This includes, for example, location, associated user roles, active users, basic process objectives, schedule requirements, regulations, operating procedures, safety regulations, dependencies on other systems, impacts, and risk analyses. Global information is shared globally and is not directly associated with any individual user. Preferably, global information is managed by a state manager.

[0054] The term "individual user information" as used herein includes an individual user's actions and / or contextual information. Individual user information is personalized and always belongs to exactly one user. Preferably, individual user information is tracked by a session manager. Contextual information preferably includes information related to the user's overall context, such as the user's role.

[0055] Preferably, the session manager and the state manager are linked due to the association between individual users and their operational roles and their participation in activities.

[0056] Preferably, user-triggered actions and requests will update individual user information, or in other words, update the state of the session manager.

[0057] Preferably, the virtual assistant handles the following steps: converting information requests into a machine-understandable format, preparing and executing relevant technical queries, processing and aggregating input results, and converting machine-generated responses into responses that can be presented to the user.

[0058] Individual user information, particularly changes within that individual user information, is provided to the state manager. This propagation of individual user information affects the session manager. In other words, individual user information is used to determine global information. For example, when a user performs a security-related task typically indicated by individual user information, that individual user information is associated with the session manager, which adjusts global information, thus alerting other users in the security-related task area of ​​another user. In other words, updates to the individual user information and the state of the session manager preferably trigger updates in the state manager, thereby triggering updates in the global information. More preferably, updates in the state manager preferably trigger updates in the session manager, thereby triggering updates to the individual user information of all relevant users.

[0059] Existing assistive systems can be operated by multiple users; however, during a single interaction, the focus is on the individual. The individual's context is sparsely utilized. The most frequently used information is personal preferences and settings, location, and identification with the relevant technological system. Contextual information such as weather or holiday seasons is rarely considered. This is because such assistants are typically unaware of the larger (social) structure and the involvement of other individuals and systems. Furthermore, since assistive systems typically focus on transferring information from a predefined database and delegating to selected predefined services, the requester's underlying motivations, completeness, satisfaction, and impact on other users are never assessed and incorporated into future actions.

[0060] Considering both global and individual user information allows for improved collaboration across different users and / or systems, while ensuring transparency, traceability, security, and consistency.

[0061] In addition, virtual assistants provide the workplace with the availability and comfort of a personal assistant.

[0062] Therefore, an improved method for controlling virtual assistants is provided.

[0063] In a preferred embodiment, determining individual user information includes tracking individual user actions and contextual information.

[0064] Therefore, an improved method for controlling virtual assistants is provided.

[0065] In a preferred embodiment, the method includes collecting user actions over a time period, analyzing the user actions to identify patterns in the user actions, using the identified patterns to determine the predicted information intent, using the predicted information intent to determine a predictive machine learning model, and using the predicted information intent to determine the predicted response.

[0066] Therefore, an improved method for controlling virtual assistants is provided.

[0067] In a preferred embodiment, the step of analyzing user actions is repeated at a predetermined frequency.

[0068] Therefore, an improved method for controlling virtual assistants is provided.

[0069] In a preferred embodiment, the method includes the steps of receiving semantic information about a desired information intent and using the received semantic information to determine the information intent.

[0070] Preferably, determining the information intent by using the received semantic information includes the user teaching the information intent to the virtual assistant.

[0071] Because virtual assistants can be addressed via a natural language interface, they sometimes fail to derive informational intent from user interactions. The user is then forced to rephrase their question or use a different input interface to obtain the desired information or submit a command. In other cases, the virtual assistant may be able to derive the user's informational intent, but the response may be insufficient or even incorrect. A typical user of such a virtual assistant is unaware of how to formulate and input new informational intents from the database of informational intents used by the virtual assistant.

[0072] Therefore, the virtual assistant knows the information intent related to creating a new information intent. When an information intent to create a new information intent is triggered by the user, the virtual assistant guides the user through a specific workflow and gathers the necessary information about the information intent, such as the name of the intent, keywords, expected action, conditions, and necessary context.

[0073] Furthermore, the virtual assistant is aware of the information intent associated with the feedback intent. This feedback intent is preferably triggered by the user with an "error" line. The feedback intent triggers a workflow within the virtual assistant to query the user for the exact location of the error and, depending on additional user input, statistically retrain the machine learning model responsible for providing the response. Alternatively, the user's feedback can be forwarded to the appropriate service.

[0074] When evaluating the quality of a machine learning model, if the model quality scores relatively low, the virtual assistant is configured to proactively ask the user for feedback. For example, the virtual assistant might ask the user, "Do you want to check valve 123?".

[0075] Preferably, the necessary information of the information intent includes the name of the information intent, typical phrases, the event the user uses to describe, important keywords, the action to be performed (e.g., performing a new analysis service or following a defined workflow), a combination of known information intents, the valid context of the information intent, and / or the response text provided by the virtual assistant.

[0076] Preferably, the feedback information intent includes voice misunderstanding, incorrect phrase intent matching, and / or errors in the results of the information retrieval service.

[0077] Preferably, the virtual assistant provides an intent editor for the user, where the user can insert desired information intents.

[0078] Therefore, the virtual assistant's ability to use new informational intents that users perceive as different behaviors is enhanced. Furthermore, users are able to provide feedback even when answers are insufficient.

[0079] According to one aspect, the virtual assistant is configured to perform the methods described herein.

[0080] Preferably, the virtual assistant includes a model scorer configured to determine whether the quality of a candidate machine learning model is acceptable.

[0081] Preferably, a computer program is provided, including instructions that, when executed by a computer, cause the computer to perform the methods described herein.

[0082] Preferably, a computer-readable data carrier is provided on which the computer program as described herein is stored. Attached Figure Description

[0083] The subject matter of the invention will be explained in more detail below with reference to preferred exemplary embodiments shown in the accompanying drawings, in which:

[0084] Figure 1 A schematic diagram of a virtual assistant is shown;

[0085] Figure 2 A schematic diagram of a method for controlling a virtual assistant is shown;

[0086] Figure 3 Another schematic diagram of a method for controlling a virtual assistant is shown;

[0087] Figure 4 Another schematic diagram of a virtual assistant is shown;

[0088] Figure 5 Another schematic diagram of a virtual assistant is shown;

[0089] Figure 6Another schematic diagram of a method for controlling a virtual assistant is shown.

[0090] The reference numerals used in the accompanying drawings and their meanings are listed in summary form in the reference numeral list. In principle, the same parts have the same reference numerals in the drawings. Detailed Implementation

[0091] Preferably, the functional modules and / or configuration mechanisms are implemented as programming software modules or processes; however, those skilled in the art will understand that the functional modules and / or configuration mechanisms can be implemented entirely or partially in hardware.

[0092] Figure 1 A virtual assistant 10 specifically designed for use in industrial workshops is illustrated. In other words, a user U, acting as an operator in the industrial workshop, uses the virtual assistant 10 to support their work. For example, user U requests a process variable value, such as the pressure of a valve, from the virtual assistant 10. The virtual assistant 10 includes an input interface 20, a control unit 30, and a model manager 40. User U inputs information request Ir into the virtual assistant 10, specifically into the input interface 20. The input interface 20 may include a graphical user interface (GUI). In this case, user U inputs their information request into the virtual assistant 10 by selecting at least one of several options provided by the GUI. Alternatively, the input interface 20 may include a language interface. In this case, user U simply formulates their information request Ir in natural language. For example, user U asks the virtual assistant 10, “How high is the pressure in valve 123?” In any case, the input interface 20 translates user U's input into a machine-understandable format that the control unit can process.

[0093] Therefore, the user's information request Ir is propagated to the control unit 30 through the input interface 20. The control unit 30 processes the information request Ir, thereby using the received information request Ir to determine the model specification Ms. In other words, the control unit 30 determines what specifications the machine learning model M should have in order to be able to determine the response R to the information request Ir.

[0094] A model specification Ms is provided to the model manager 40. The model manager 40 uses the provided model specification Ms to determine a suitable machine learning model M. In other words, the model manager 40 determines suitable parameterization for the machine learning model M. In the provided exemplary case, the model manager 40 provides a machine learning model M that is capable of providing a response R to an information request Ir regarding the pressure of a specific valve.

[0095] Therefore, the determined machine learning model M is provided to the control unit 30. The control unit 30 uses the machine learning model M to determine the response R to the information request Ir. The control unit 30 thus determines the necessary inputs to the machine learning model M, specifically mapping the equipment in the industrial plant to the corresponding sensors using the topology information of the industrial plant. In other words, the control unit 30 maps the sensor outputs of relevant real-world sensors to the corresponding inputs to the machine learning model M.

[0096] The response R is then provided to input interface 20, which is also configured to convert the response R into a natural language format. For example, virtual assistant 10 is configured to use input interface 20 to provide a direct response R to user U in the same language that user U has already used to input the information request Ir.

[0097] Figure 2 A schematic diagram of a method for controlling the virtual assistant 10 is shown.

[0098] In the first step S10, the input interface receives an information request Ir, wherein the information request Ir includes at least one request for receiving information about at least a portion of the industrial workshop. In the second step S20, the control unit 30 uses the received information request Ir to receive a model specification Ms. In the third step S30, the model manager 40 uses the model specification Ms to determine a machine learning model M. In the fourth step S40, the control unit 30 uses the determined machine learning model M to provide a response R to the information request Ir.

[0099] Figure 3 Another schematic diagram shows the control method of the virtual assistant 10.

[0100] and Figure 2 Conversely, in another step S50, the control unit 30 uses the received information request Ir to identify the information intent I. Based on the information intent I, the model specification Ms is determined. The information intent I provides an additional layer of depth for the determination of the model specification Ms.

[0101] Furthermore, step S30, determining the machine learning model M, is further specified. In another step S60, the model database is checked to find a suitable machine learning model M that can be used to determine the response R to the information request Ir. If a suitable machine learning model M according to the model specification Ms is found in the model database, the machine learning model M is provided to the control unit 30. As described above, in step S40, the control unit 30 uses the determined machine learning model M to provide the response R to the information request Ir.

[0102] If it is determined that no suitable machine learning model M according to the model specification Ms does not exist in the model database, then in another step S70, the user U is notified of the delay. Simultaneously, in another step S80, the automatic machine learning process autoML is initiated to generate a suitable machine learning model M according to the model specification Ms. This potential machine learning model is referred to as the machine learning model candidate Mc.

[0103] In another step S90, the model quality of the machine learning model candidate Mc is tested. If the model quality of the machine learning model candidate Mc is suitable, and if in another step S100 it is determined that the machine learning model candidate Mc has acceptable model quality, the method jumps to another step S110, where the machine learning model candidate Mc is stored as another machine learning model M in the model database. The method then jumps to step S40, where the control unit 30 provides a response R to the information request Ir using the determined machine learning model M.

[0104] If, in another step S100, it is determined that the machine learning model candidate Mc does not have acceptable model quality, the method jumps to another step S120, where the user U is notified of the unsuccessful generation of the machine learning model, thus the method fails. Therefore, the user U is notified that he cannot expect a response R to his information request Ir. Thus, the user U must either accept that the virtual assistant 10 cannot provide a response R, or must retry another information request Ir. This alternative information request Ir can be made possible by rewriting the initial information request Ir.

[0105] Figure 4 Another schematic diagram of the virtual assistant 10 is shown. (Compared to...) Figure 1 Compared to the virtual assistant 10, the model manager 40 is extended by a self-organizing model generator 50 and an automated machine learning (autoML) pipeline 60. The autoML pipeline 60 includes a data selector 61, a model selector 62, a model generator 63, and a model evaluator 64.

[0106] If no suitable machine learning model M is available in the model database, the self-organizing model generator 50 determines a new machine learning model M. The self-organizing model generator 50 determines the new machine learning model M in real time.

[0107] The generation of the machine learning model M is supported by the autoML pipeline. Based on the model specification Ms, data selector 61, model selector 62, and model generator 63 determine the machine learning model M. The machine learning model M is then evaluated by model evaluator 64 to ensure that the machine learning model M is suitable for determining the response R to the information request Ir. The determined machine learning model M is propagated from the autoML pipeline through model generator 50 and model manager 40 to control unit 30.

[0108] Figure 5 Another schematic diagram of the virtual assistant 10 is shown. Compared to the already described virtual assistant 10, the control unit 30 is extended by a session manager 31 and a state manager 32. The session manager 31 tracks the actions and context information of individual users U. The information associated with user U through the session manager 31 is personalized and always belongs to exactly one user U. This information is referred to as individual user information I1. The state manager 32 manages the user's ongoing processes, workflows, tasks, and actions, referred to as global information Ig. This includes, for example, the user's location, the roles associated with the user, active users, basic goals, progress, requirements regulations, operating procedures, safety regulations, dependencies on other systems, impacts, and risk analysis. This information is globally shared and is not associated with a single user U.

[0109] Session Manager 31 receives user action information Ia. For example, it receives user action information Ia through input interface 20. However, it can also receive user action information Ia through another channel. Session Manager 31 uses the received user action information Ia to determine individual user information Iu. In other words, Session Manager 31 associates specific user action information Iu with a specific user U. Individual user information Iu is provided to State Manager 32. State Manager 32 uses individual user information Iu to execute an overview request, requesting service overview information I1 from Service Overview 70. Service Overview 70 relates to the structure of the industrial workshop in the segment associated with user U. In other words, Service Overview 70 is an indication of all entities in the industrial workshop associated with user U, their interconnections, and their status. State Manager 32 uses the received service overview information I1 to determine global information Ig. Therefore, Session Manager 31 uses global information Ig to determine the user's response R.

[0110] In an example of a hazardous operation, user U initiates a task at a specific location and reports that location via virtual assistant 10. Session manager 31 updates user U's state and propagates the corresponding individual user information Iu to state manager 32. State manager 32 receives individual user information Iu, including user U's location and the start of user activity. State manager 32 checks for potential hazards at that location and activates a monitor to continue updating the location's state. Furthermore, state manager 32 requests a report from the global monitor in service overview 70. In this case, service overview information I1 in service overview 70 reports an anomaly that conflicts with human (i.e., user U) safety regulations. The monitoring task in state manager 32 reacts to the input service overview information I1 and identifies all relevant users within the given location. Additionally, state manager 32 gathers criteria to react when available and issues an alert to session manager 31. This alert is a portion of the global information Ig provided by state manager 32 to session manager 31. Session manager 31 receives the alert and proactively informs user U about their current location. For example, if a task ends or user U moves to a different location, this will implicitly terminate monitoring of the starting location, thus avoiding unnecessary notifications. For instance, if user U has already notified virtual assistant 10 of their intention to start a task ahead of schedule, the virtual assistant can check that the location where the action began has actually started.

[0111] In another example of a hazardous operation, user U intends to begin a task at a specific location. En route to that location, an anomaly is observed. User U issues an information request Ir to virtual assistant 10 regarding the status of user U's current location. Session manager 31 updates the user's status and propagates the individual user information to state manager 32. State manager 32 requests to locate the service overview information I1 for the requested location in the global monitor. This request does not return an anomaly status. The user requests confirmation from another person in the control room. Visual and audio information of the current situation, along with a description of the situation, is bundled and forwarded as individual user information Iu. State manager 32 provides the operator in the control room with the information request Ir containing individual user information Iu. The situation is manually assessed, and global information Ig is determined for state manager 32 based on the manual assessment. In another example, if multiple operators can attest to the request, they will all be notified. Once an operator actively works on the task, notifications to other operators are removed or marked as ongoing work. In cases of security risks, virtual assistant 10 also notifies users in the relevant area. The original requester will receive only one notification because the virtual assistant 10, with the help of the session manager 32, confirms that the response and the alarm belong to the same event.

[0112] Figure 6Another schematic diagram of a method for controlling a virtual assistant is shown. In step Z10, user actions of user U are collected over a predetermined time period. In another step Z20, the user actions are analyzed to identify patterns in the user actions. In another step Z30, it is determined whether a pattern of user actions has been identified. If no pattern has been identified, the method returns to step Z20. If a pattern has been identified, the relevant information intent I is associated with the identified pattern. In other words, it is determined which information intent I relates to the identified pattern. In another step Z40, the information intent I and the associated identified pattern are stored in an intent database. Thus, if the stored pattern is identified in a future information request, the associated information intent I can be predicted. In this way, shortcuts to information intent I can be learned, or the user's next action can be suggested based on the identified pattern.

[0113] List of reference numerals

[0114] 10 Virtual Assistants

[0115] 20 Input Interfaces

[0116] 30 Control Unit

[0117] 31 Session Manager

[0118] 32 State Manager

[0119] 40 Model Manager

[0120] 50 Self-organizing model generators

[0121] 60autoML production line

[0122] 61 Data Selector

[0123] 62 Model Selector

[0124] 63 Model Generator

[0125] 64 Model Evaluator

[0126] 70 Service Overview

[0127] Ir Information Request

[0128] Ms Model Specification

[0129] M Machine Learning Model

[0130] R response

[0131] Mr. Model Request

[0132] U User

[0133] Iu Individual User Information

[0134] Ig Global Information

[0135] Ia User Action Information

[0136] Rl Overview Request

[0137] I1 Service Overview Information

[0138] I. Information Intent

[0139] Mc Machine Learning Model Candidates

[0140] S10 Receives Information Request

[0141] S20 Determines Model Specifications

[0142] S30 Determine the machine learning model

[0143] S40 provides responses to information requests.

[0144] S50 Identification Information Intent

[0145] S60 checks the model database used for suitable machine learning models.

[0146] S70 notifies users about delays

[0147] S80 Generates Machine Learning Model Candidates

[0148] S90 Test Machine Learning Model Candidate

[0149] The S100 model quality is acceptable.

[0150] S110 storage machine learning model candidate

[0151] S120 notified the user of the failure.

[0152] Z10 collects user actions

[0153] Z20 analyzes user actions

[0154] Z30 determines whether the mode is recognized.

[0155] Z40 Storage Mode and Information Intent

Claims

1. A method for controlling a virtual assistant (10) for an industrial plant, comprising: receiving, by an input interface (20), an information request (Ir), wherein the information request (Ir) comprises at least one request for receiving information about at least a part of the industrial plant; determining, by a control unit (30), a model specification (Ms) using the received information request (Ir), the model specification indicating technical requirements for a machine learning model; determining, by a model manager (40), the machine learning model (M) using the model specification (Ms) and providing the determined machine learning model to the control unit; determining, by a session manager (31), user action information (Ia) related to tracked individual user's actions and context information; determining, by the session manager (31), individual user information (Iu) using the user action information (Ia), wherein determining the individual user information (Iu) comprises tracking the user's actions and context information; requesting, by a state manager (32), service overview information (Il) from a service overview (70) based on the individual user information (Iu), wherein the service overview (70) relates to a structure of the industrial plant in a section related to the user (U); determining, by the state manager (32), global information (Ig) using the service overview information (Il), wherein the global information (Ig) comprises information about ongoing processes, workflows, tasks and actions of the industrial plant; and determining, by the session manager (31), a response (R) for the user (U) using the global information (Ig); providing, by the control unit (30), the response (R) to the information request (Ir) using the determined machine learning model (M).

2. The method according to claim 1, comprising: identifying, by the control unit (30), an information intent (I) using the received information request (Ir); and determining the model specification (Ms) using the information intent (I).

3. The method according to claim 1, wherein determining a machine learning model (M) comprises: checking, using the model specification (Ms), whether a suitable machine learning model is stored in a model database.

4. The method according to claim 3, wherein if it is determined that a suitable machine learning model is stored in the model database, the method comprises the step of: determining the response (R) to the information request (Ir) by using the stored machine learning model.

5. The method according to claim 2, wherein determining the response to the information request (Ir) comprises: providing, by the control unit (40), a model input, wherein the model input is determined by using the information intent (I); determining the response (R) by inputting the model input to the machine learning model (M).

6. The method according to claim 3, ​ wherein if it is determined that no suitable machine learning model is stored in the model database, the method comprises the steps of: providing a delayed response to the user (U).

7. The method of claim 2, comprising: determining a machine learning model candidate using the information intent (I) by the autoML pipeline (60); testing a model quality of the machine learning model candidate; if the model quality is acceptable, determining the machine learning model (M) using the machine learning model candidate; and if the model quality is not acceptable, informing the user (U) that model generation was not successful.

8. The method of claim 2, identifying the information intent comprises: wherein converting the information request into a machine understandable format.

9. The method of any of claims 1-3, providing a response (R) comprises: wherein converting the determined response into a user understandable format.

10. The method of any of claims 1-3, comprising: collecting user actions over a time period; analyzing the user actions, thereby identifying a pattern of user actions; associating an information intent (I) with the identified pattern; and using the identified pattern to predict the information intent (I) of an information request.

11. The method of claim 10, wherein the step of analyzing the user actions is repeated at a predetermined frequency.

12. The method of any of claims 1-3, comprising: receiving semantic information for a desired information intent; determining the information intent using the received semantic information.

13. A virtual assistant configured to perform the method of any of claims 1-12. ​

Images