Method and system for operating a technical device with a controller

Abstract

The invention relates to a computer-implemented method for operating a technical device with a controller, comprising the following steps: a) providing a user request relating to the operation of the technical device, b) providing a trained operating model on the basis of artificial intelligence for operating the technical device, c) forming a search request from the user request, and determining a first response to the search request by means of a connected "retrieval augmented generation", RAG, system formed by a connected first computing device, d) forming an estimation request from the user request, and determining a second response to the estimation request by a connected "federated learning", FL, system for operating the technical device, which FL system is formed by a connected second computing device, e) combining the first response and the second response with the aid of a provided weighting function, and generating an adapted demand for input for the operating model with the aid of the weighting function, and determining, with the aid of the operating model, a response to the adapted demand for input by means of a connected third computing device; f) operating the technical device by means of the controller with the aid of control signals which are derived from the third computing device from the previously obtained response.

Description

[0001] 202500701

[0002] 1

[0003] Method and system for operating a technical device

[0004] with a control device

[0005] The invention relates to a computer-implemented method and a system for operating a technical device with a control device.

[0006] The invention further relates to a computer program product.

[0007] In industry, as part of the digitalization process, machine learning-based models are increasingly used in the manufacturing or processing of products, as well as in the operation of technical systems and equipment, to control, operate or monitor a technical device using a control device or a monitoring device and applying appropriate control instructions, whereby the use of a model with high accuracy, i.e., a representation of the respective physical representation of the model that is as accurate as possible, plays an important role.

[0008] Adapting "Industrial Foundation" (IFM) models based on prompt tuning (PT), Retrieval-Augmented Generation (RAG), or fine-tuning involves changing the behavior of a pre-trained model to better fit a domain.

[0009] PT and RAG are lightweight approaches to adapt IFMs to a specific context without retraining, which is essential to support versatile applications in industry.

[0010] Retrieval-augmented generation (RAG) refers to a software system that combines information retrieval with a large language model (LLM). A query submitted to the system can access information from (external) sources, databases, or the World Wide Web, instead of just the model's training data. This increases the accuracy and robustness of the generated content by providing the models with current and specific information. Retrieval-augmented generation (RAG) essentially consists of four stages: data indexing, data retrieval, augmentation, and response generation.

[0011] Federal learning (FL) describes a machine learning technique in which a model is trained on multiple devices. Each participating device has its own local dataset that is not shared with other participants. In contrast, traditional machine learning uses a central dataset.

[0012] 2

[0013] This technique allows multiple participants to build a shared robust model, taking into account critical issues such as data protection, data security, access rights, and access to heterogeneous data.

[0014] Prompt tuning involves modifying the prompt or input format assigned to the model to control its behavior for a specific task.

[0015] This can include creating customized prompts or instructions that cause the model to achieve the desired results.

[0016] It is advantageous to use a simple approach that requires minimal computing resources compared to retraining the entire model, which can be particularly effective for tasks where the input format strongly influences the model's behavior.

[0017] However, the method may not be suitable for tasks requiring significant model fitting or fine-grained control, and may require specialized expertise to design effective prompts.

[0018] The customization of the input prompt is crucial for quickly adapting a pre-trained model to new tasks or domains without extensive retraining. This allows users to leverage existing model capabilities while simultaneously providing task-specific guidance.

[0019] In Retrieval-Augmented Generation (RAG), a generative model is combined with a "retriever," allowing the model to access external knowledge during generation. By retrieving relevant information from a large knowledge database, the model can produce more accurate and contextually relevant outputs. The advantages include improved contextual understanding through the incorporation of external knowledge, enhanced quality and relevance of the generated results, particularly for tasks requiring factual accuracy or contextual understanding, and a flexible approach to utilizing external data without altering the model architecture.

[0020] However, the retrieval process requires access to a large knowledge base, which may not always be available or feasible. Accessing external knowledge during generation can lead to undesirable latency, which can negatively impact real-time applications.

[0021] 3

[0022] It is advantageous when queries rely on contextual understanding and factual accuracy. This allows the model to seamlessly access and integrate external knowledge, improving performance in tasks such as question answering, summarizing, and dialogue generation.

[0023] RAG is a beneficial factor for the success of IFM implementation.

[0024] In current technology, data collection and aggregation, for example by collecting relevant data from various internal sources such as documents, reports, emails, customer interactions and databases, is often a complex and time-consuming process.

[0025] Ontology development and knowledge representation, namely the design of an effective ontology or knowledge graph that captures domain-specific concepts, relationships, and hierarchies, is a non-trivial task.

[0026] Developing ontologies requires expertise and collaboration among subject matter experts to define relevant entities, attributes, and semantic relationships. Creating a well-structured ontology facilitates efficient data organization, retrieval, and knowledge inference within the knowledge base.

[0027] The biggest challenge, however, is maintaining and updating the knowledge: Maintaining and updating the internal knowledge database is an ongoing effort that requires dedicated resources and processes.

[0028] The object of the invention is to provide a solution that solves a state problem, namely requiring reliable access to a large knowledge database for each retrieval with improved latency, which may not always be available or feasible.

[0029] Thus, a new solution should improve availability, system integrity and accuracy while remaining state-of-the-art.

[0030] The problem according to the invention is solved by a computer-implemented method for operating a technical device with a control device, comprising the following steps:

[0031] a) Providing a user request regarding the operation of the technical device, b) Providing a trained operating model based on artificial intelligence for the operation of the technical device, 202500701

[0032] 4

[0033] c) Forming a search query from the user query and determining a first answer to the search query by a connected Retrieval-Augmented Generation (RAG) system formed by a connected first computing device; d) Forming an estimation query from the user query and determining a second answer to the estimation query by a connected Federated Learning (FL) system for operating the technical device, formed by a connected second computing device.

[0034] e) Combining the first answer with the second answer using a provided weighting function, and generating a customized prompt for the operating model using the weighting function, and determining a response to the customized prompt using the operating model, by means of an associated third computing device,

[0035] f) Operation of the technical device by the control device using control signals which are derived from the previously received response by the third computing device.

[0036] The Mixture of Experts (MoE) approach according to the invention integrates the models “retrieval-augmented generation” (RAG) and federated learning (FL) to utilize their respective strengths and, using this aggregation, generates a customized input request to a large language model, thereby enabling a reliable and accurate response to a user query to be provided in a simple manner.

[0037] In MoE, a so-called "gating" network dynamically determines the contribution of each model (RAG and FL) based on the input query or the context.

[0038] The RAG expert uses external knowledge from a large knowledge database to generate context-relevant answers.

[0039] Meanwhile, the FL expert learns from decentralized data sources while maintaining data privacy through collaborative training, which can improve system integrity.

[0040] The results from the RAG and FL experts are combined using the gating coefficients generated by the gating network, allowing the MoE model to adaptively adjust each expert's contribution—that is, flexibly, depending on the operational situation or context of the technical device or client—and also over time. This hybrid approach improves the model's performance by leveraging both RAG's contextual understanding and FL's collaborative learning capabilities.

[0041] 5

[0042] Overall, MoE enables more versatile and effective solutions for tasks that require contextual understanding and collaboration across decentralized data sources.

[0043] The invention achieves a hybrid integration of RAG and FL.

[0044] In contrast to traditional approaches, which typically focus on either "retrieval-augmented generation" or "federated learning" independently, the MoE approach now integrates both models, resulting in overall improved performance by simultaneously leveraging the contextual understanding of RAG and the collaborative learning of FL.

[0045] Furthermore, this allows for an adaptive model contribution, as the contribution of each expert (RAG and FL) is dynamically adjusted via a gating network based on the input query or context.

[0046] This adaptive mechanism ensures that the model can flexibly leverage the strengths of each expert to optimize performance for different scenarios and data distributions. Furthermore, it improves contextual understanding and collaboration, as both can be necessary across decentralized data sources.

[0047] By combining RAG's ability to incorporate external knowledge with FL's privacy-friendly collaborative learning, MoE offers a unique solution that addresses both the challenges of contextual understanding and data decentralization.

[0048] Furthermore, the collaboration takes place while respecting privacy, which improves data security, as FL's privacy-friendly features are cleverly used to ensure that sensitive data remains protected while enabling collaborative training across decentralized devices or data sources.

[0049] Improving data privacy also improves system integrity and protection against data theft or manipulation, and consequently, the availability of the system.

[0050] The model – the operating model – is used to govern the operation of a technical device connected to a corresponding control device. The resources of the device platform for operating the technical device, in terms of software and hardware, may include, for example, the computing power, storage capacity, and / or communication capacity of the computing device. 202500701

[0051] 6

[0052] Typically, a control device has limited resources for processing models, such as a low-performance processor and limited memory, compared to the resources of a workstation or a server in a cloud infrastructure. The technical device can vary widely, for example, a pump, an electronic controller such as an inverter, an electronic sensor such as an accelerometer or imaging sensor such as a camera, a production machine such as an assembly robot, a transport robot for logistics, a CNC milling machine, etc.

[0053] The technical device can map its operation through operating parameters, which are recorded, for example, by appropriate sensors, or determined by analyzing communication messages during operation.

[0054] Operating parameters can include, for example, current consumption, voltage values ​​of subsystems, or power output of the technical device, but also heat, noise or vibration emissions, or the number or size of operating data sets generated by the technical device.

[0055] The control device for such a technical device can be an analog electronic circuit, a pneumatic control, or a microprocessor circuit with digital control connections, which in turn controls power electronic components to operate actuators.

[0056] The control device generates control instructions or control signals for the operation of the connected technical device, which are generated using an artificial intelligence-based model, namely the application model generated from a provided large IFM / LVM.

[0057] The criteria for executing the model using control instructions or signals on a computing device such as an edge device can be performance criteria of the computing device, describing the efficiency of processing computing, storage, and / or communication properties, or performance criteria for the control device in generating and executing corresponding control signals for operating the technical device. Using a customized model for operating the technical device is particularly advantageous when, for example, several production or processing machines are operated within a client-server system, but these machines manufacture different products and therefore require different AI models for the respective manufacturing / processing processes of those products.

[0058] 7

[0059] Similarly, technical devices such as pumps or motors can also be operated with different performance levels by different clients using different AI models.

[0060] The prompt refers to a provided, pre-trained large language model (LLM).

[0061] In a further development of the invention, it is provided that the RAG system comprises a database and / or a pre-trained model based on artificial intelligence for the operation of the technical device, each of which includes a specific set of data describing the operation of the technical device in its operating environment and / or application, wherein the operating environment describes the physical arrangement and / or physical operating parameters of the technical device.

[0062] The specific data set describes, for example, the operation of the technical device in its operating environment, such as the current operating temperature, the limits of the permissible operating temperature, and the permissible or impermissible duration outside these limits. The specific data set also describes, for example, the application of the technical device, such as the use of a pump in an air conditioning system or a wastewater treatment plant, or the use of an inverter for a specific power application, such as in a photovoltaic system or the operation of a motor controller.

[0063] In a further development of the invention, it is provided that the method for operating the technical device is carried out repeatedly, and that respective physical operating parameters for the operation of the technical device are recorded, and that predetermined value ranges are defined for the operating parameters, for which respective weight functions are defined, and in step e) a weight function corresponding to the current operating parameter is applied.

[0064] In a further development of the invention, it is provided that the weight function is defined for a respective combination of values ​​or value ranges of the individual physical operating parameters, and the weight function is represented by a weight model based on artificial intelligence, which is used to select the weight function.

[0065] The operating parameters relate to the operation of the connected technical device. 202500701

[0066] 8

[0067] A currently provided weight function is therefore formed from a provided weight model and determined or selected based on combinations of values ​​or value ranges of the individual physical operating parameters.

[0068] The weight model is based on artificial intelligence and is manually trained using historical data for such combinations of operating parameters, for example, particularly favorable or particularly unfavorable operating conditions or operating modes, and marked accordingly.

[0069] In other words, this can be achieved, for example, by using a separate weight model based on artificial intelligence, in which several weight functions form training data, and the weight model makes a prediction for a corresponding favorable aggregation.

[0070] In a further development of the invention, it is provided that the physical operating parameters are electrical operating parameters and / or mechanical operating parameters and / or thermal operating parameters and / or meteorological operating parameters.

[0071] The operating parameters can be either static or dynamic.

[0072] Electrical operating parameters can include, for example, current consumption, power output, the level or fluctuation of the operating voltage.

[0073] Dynamic mechanical operating parameters can include, for example, vibrations of a pump, which are detected using appropriate sensors, or a generated torque of a motor.

[0074] Static mechanical operating parameters can include, for example, the installation position or the arrangement in relation to other devices or subsystems.

[0075] Thermal operating parameters can include, for example, heat emissions or heat distribution on the technical device or its components, such as a semiconductor component.

[0076] Meteorological operating parameters can include the humidity or the ambient temperature of the technical device.

[0077] In a further development of the invention, the FL system comprises a pre-trained artificial intelligence-based model for operating the technical device, which is provided on a client-server system with a server and at least one client, and is generated from a global model on the server and a respective local model on the at least one client.

[0078] 9

[0079] This efficiently protects the data used to operate the technical device, thereby improving data processing efficiency and communication efficiency in the client-server system.

[0080] In a further development of the invention, it is provided that the control signals contain control instructions for execution by a processor.

[0081] This makes it easy to use the obtained model to control or monitor the operation of the technical device.

[0082] In a further development of the invention, it is provided that the control device is included by a client of a client-server system, and is preferably an edge device.

[0083] This ensures that the process can be used in a particularly advantageous environment, resulting in low system complexity and efficient operation of the technical equipment.

[0084] In a further development of the invention, it is provided that the first, the second and the third calculating device are formed by a common calculating device.

[0085] This allows for low system complexity and efficient operation of the technical equipment.

[0086] The training data is preferably operational data from the operation of the technical device, for example, recorded by sensors such as accelerometers, which monitor the operation of a motor.

[0087] User requests are primarily for operational analysis tasks, such as predictive maintenance planning for the technical equipment, whereby the searches can be performed by a user or automatically by a monitoring device.

[0088] The control signals are preferably control instructions intended for execution by a processor, for example program code.

[0089] In a further development of the invention, it is provided that the control device is included by a client of a client-server system, and is preferably an edge device.

[0090] This allows the application model to be executed efficiently on a control device of a client-server system, and thus, starting from a central LVM / IFM, a reduced or compressed application model can be easily distributed to multiple clients, and furthermore, an individual / specific application model can be provided for each connected client.

[0091] 10

[0092] Therefore, a particularly accurate application model can be generated within the system and applied to a specific technical device.

[0093] The problem according to the invention is solved by a system for operating a technical device with a connected control unit, further comprising a processor and a memory, wherein the system is configured to carry out the method according to one of the preceding claims.

[0094] In a further development of the invention, it is provided that the system is a client-server system, and the control device is comprised of a client of the client-server system, and is preferably an edge device.

[0095] The problem according to the invention is solved by a computer program product comprising instructions which, when the program is executed by a computer, cause it to execute the method / steps of the method according to one of the preceding method claims.

[0096] The computer is encompassed by the system.

[0097] The invention is described in more detail in the following figures using an exemplary embodiment. The figures show in

[0098] Fig. 1 shows an embodiment of the invention in the form of a flowchart,

[0099] Fig. 2 shows an embodiment of the invention in the form of a pseudocode,

[0100] Fig. 3 shows a further embodiment of the invention in the form of a flowchart, Fig. 4 shows an embodiment of the invention in the form of a system.

[0101] Fig. 1 shows an embodiment of the invention in the form of a flowchart.

[0102] A provided large dataset (LDS) is mapped into a base large language model (B-LLM) by a pre-training (PT).

[0103] A request Q from a user U in the form of an input prompt sends an input to the base model B-LLM using a request / response system QAS, which generates a corresponding response RES.

[0104] The QAS request / response system first initiates the request Q to a

[0105] Subsystem CMoE, which comprises a combination of "Mixture of Experts" and forms a so-called "gating network", continues as search SE. 202500701

[0106] 11

[0107] The gating network uses the input query or its context and generates so-called gating coefficients, which represent the importance of each expert (RAG and FL) for this input.

[0108] The gating coefficients are usually normalized to sum to one to ensure that they represent a valid probability distribution.

[0109] Furthermore, an estimate P is calculated from the subsystem CMoE using a federated learning model FLM.

[0110] Expert networks for RAG and FL are represented by respective expert networks, namely a corresponding RAG or FL model.

[0111] The RAG expert generates a response based on the input query, using the retrieval-augmented generation approach, creating a specific set of SDS data, for example depending on the organization or domain.

[0112] The FL expert generates a prediction based on the input query using the Federated Learning model FLM.

[0113] A customized input prompt (AP) is generated from the two expert networks for RAG and FL using a so-called mixing layer. This layer combines the results of the RAG and FL experts using the gating coefficients generated by the gating network, which are weights for the individual sub-results. This is typically done by calculating a weighted sum of the outputs, with the weights determined by the gating coefficients.

[0114] Subsequently, a training process takes place, during which the gating network and the expert networks (RAG and FL) are trained together using backpropagation and a corresponding loss function.

[0115] This weighting model learns to assign appropriate weights to the RAG and FL experts based on the input query or context, thereby optimizing the overall performance of the system.

[0116] The weight model outputs a corresponding response to the adapted input prompt AP, which is returned to the user using an optionally further processed response R and includes, for example, a statement about the maintenance of the technical device TD1 predicted by the model.

[0117] Fig. 2 shows an embodiment of the invention in the form of pseudocode for the method according to the invention, wherein a definition of a simple "gating" network is described.

[0118] 12

[0119] First, an example of a simple linear assignment with a subsequent softmax is shown.

[0120] In mathematics, the so-called softmax function or normalized exponential function is a generalization of the logistic function that transforms a K-dimensional vector z with real components into a K-dimensional vector o(z), also as a vector of real components, into the range (0,1), where the components sum to 1.

[0121] Then an expert network is defined for RAG and FL, where, for example, the RAG expert always estimates one, and the FL expert estimates based on an input characteristic.

[0122] Furthermore, a definition of a "Mixture of Experts" function is given, in which the gating coefficients / weights are assigned values, expert estimates are also assigned, and a combined output is provided using the gating weights.

[0123] Furthermore, an example of input characteristics is given.

[0124] Finally, it is shown how a combined output can be obtained using "Mixture of Experts".

[0125] Fig. 3 shows a further embodiment of the invention in the form of a flowchart. The method for operating a respective technical device TD1-TD3 with a respective control device on the client C1-C3 can be computer-implemented, at least in individual steps.

[0126] The following steps are included:

[0127] a) Providing a user request Q regarding the operation of the technical device TD1,

[0128] b) Providing a trained artificial intelligence-based model B-LLM for operating the technical device TD1,

[0129] c) Forming a search query S from the user query Q, and determining a first response to the search query S by a connected "Retrieval-Augmented Generation" (RAG) system formed by a connected first computing device,

[0130] d) Forming an estimation request P from the user request Q, and determining a second response to the estimation request P through a connected “federated202500701

[0131] 13

[0132] Learning", FL-, system for operating the technical device TD1, which is formed by a connected second computing device,

[0133] e) Combining the first answer with the second answer using a provided weighting function, and generating a customized prompt AP for the B-LLM model using the weighting function, and determining a response RES to the customized prompt using the B-LLM model, by means of an associated third computing device,

[0134] f) Operation of the technical device TD1 by the control device using control signals CSIG, which are derived by the third computing device from the previously received response RES.

[0135] The user request Q can be made either manually by a user U or automatically by a corresponding operational monitoring device.

[0136] The RAG system preferably includes a database and / or a pre-trained artificial intelligence model for operating the technical device TD1, each of which includes a specific data set SDS.

[0137] The specific data set SDS describes the operation of the technical device TD1 in its operating environment and / or application.

[0138] The operating environment describes the physical arrangement and / or physical operating parameters of the technical device.

[0139] The procedure for operating the technical device TD1 can be repeated. During this process, the respective physical operating parameters for the operation of the technical device TD1 are recorded.

[0140] Predefined value ranges are set for the operating parameters.

[0141] Specific weight functions are defined for each operating parameter.

[0142] In step e), a weight function corresponding to the current operating parameter is applied, which can be dynamically determined or selected.

[0143] It is advantageous if respective weight functions are defined, which are determined based on combinations of values ​​or value ranges of the individual physical operating parameters.

[0144] The weight functions can be represented by a weight model based on artificial intelligence, which is used to select the current weight function. 202500701

[0145] 14

[0146] The physical operating parameters can be electrical operating parameters and / or mechanical operating parameters and / or thermal operating parameters and / or meteorological operating parameters.

[0147] In this example, the FL system comprises a pre-trained artificial intelligence-based model for operating the technical device TD1, whereby the model training is carried out using operational data and parameters from the operation of the technical device TD1, for example with permissible and / or impermissible value ranges for these operational data and parameters, and optionally by applying corresponding labels.

[0148] This model is deployed on a client-server system with one server and at least one client.

[0149] The model is generated from a global model on the server and a respective local model on at least one client.

[0150] The CSIG control signals preferably contain control instructions for execution by a processor.

[0151] The control device is included in a client of a client-server system and is preferably an edge device.

[0152] The first, second and third computing devices can communicate with each other and are preferably formed from a common computing device.

[0153] The generated input prompt B-LLM refers to a provided, pre-trained large language model.

[0154] Fig. 4 shows a system as an embodiment of the invention in the form of a client-server system on which the method according to the invention can be carried out.

[0155] The client-server system has a server SE and clients C1-C3, which are connected to respective technical devices TD1-TD3 and are controlled with control instructions via corresponding control signals CSIG.

[0156] For example, a user request concerning the operation of the technical device can be made by user U to client C1.

[0157] Accordingly, the client-server system comprises a processor and memory to execute the method according to the invention, wherein parts of the method are preferably executed on the high-performance server SE. 202500701

[0158] 15

[0159] The control device can be included by a respective client C1-C3 of the client-server system and is preferably an edge device, particularly preferably an edge device with a computing device in the form of an ASIC or FPGA semiconductor component.

[0160] Regardless of the grammatical gender of a given term, persons of male, female, or other gender identities are included. 202500701

[0161] Reference symbol list

[0162] AP-adapted prompt

[0163] B-LLM Base of Large Language Model (C1-C3 Client)

[0164] CMoE combination “Mixture of Experts”

[0165] CSIG control signal, control instructions

[0166] FLM federated learning model

[0167] LDS Large Dataset

[0168] P Estimate Request

[0169] PT pre-training

[0170] Q User Inquiry

[0171] QAS Request / Response System

[0172] R, RES answer

[0173] S search query

[0174] SDS-specific data set, by organization / domain

[0175] SE Server

[0176] TD1-TD3 technical device

[0177] U users

Claims

202500701 17 Patent claims 1. Computer-implemented method for operating a technical device (TD1) with a control device, comprising the following steps: a) Providing a user request (Q) regarding the operation of the technical device (TD1), b) Providing a trained operating model based on artificial intelligence (B-LLM) for the operation of the technical device (TD1), c) Forming a search query (S) from the user query (Q), and determining a first response to the search query (S) by a connected "Retrieval-Augmented Generation" (RAG) system formed by a connected first computing device, d) Forming an estimation request (P) from the user request (Q), and determining a second response to the estimation request (P) by a connected “federated learning” (FL) system for operating the technical device (TD1), which is formed by a connected second computing device, e) Combining the first answer with the second answer using a provided weighting function, and generating a customized prompt (AP) for the operating model (B-LLM) using the weighting function, and determining a response (RES) to the customized prompt (AP) using the operating model (B-LLM) by a connected third computing device, f) Operation of the technical device (TD1) by the control device using control signals (CSIG) which are derived by the third computing device from the previously received response (RES).

2. Method according to the preceding claim, wherein the RAG system comprises a database and / or a pre-trained artificial intelligence model for operating the technical device (TD1), each of which includes a specific data set (SDS) describing the operation of the technical device (TD1) in its operating environment and / or application, wherein the operating environment describes the physical arrangement and / or physical operating parameters of the technical device.

3. Method according to one of the preceding claims, wherein the method for operating the technical device (TD1) is repeatedly carried out and respective physical operating parameters for the operation of the technical device (TD1) are recorded, and predetermined value ranges are defined for the operating parameters, for which respective 202500701 18 Weight functions are defined, and in step e) a weight function corresponding to the current operating parameter is applied.

4. Method according to the preceding claim, wherein the weight function is defined for a respective combination of values ​​or ranges of values ​​of the individual physical operating parameters, and the weight function is represented by a weight model based on artificial intelligence, which is used to select the weight function.

5. Method according to any of the preceding claims, wherein the physical operating parameters are electrical operating parameters and / or mechanical operating parameters and / or thermal operating parameters and / or meteorological operating parameters.

6. Method according to one of the preceding claims, wherein the FL system comprises a pre-trained artificial intelligence-based model for operating the technical device (TD1), which is provided on a client-server system with a server and at least one client, and is generated from a global model on the server and a respective local model on the at least one client.

7. Method according to any of the preceding claims, wherein the control signals (CSIG) comprise control instructions for execution by a processor.

8. Method according to any of the preceding claims, wherein the control device is comprised of a client of a client-server system, and is preferably an edge device.

9. Method according to one of the preceding claims, wherein the first, the second and the third calculating device are formed by a common calculating device.

10. System for operating a technical device with a connected control unit, further comprising a processor and a memory, wherein the system is configured to perform the method according to one of the preceding claims.

11. System according to the preceding claim, wherein the system is a client-server system, and the control device is comprised of a client (C1) of the client-server system, and is preferably an edge device.

12. Computer program product, comprising instructions which, when the program is executed by a computer, cause it to perform the procedure(s) of the procedure according to any of the preceding procedure claims.

Images