Method for operating an industrial environment comprising at least two different systems using an electronic computing device, computer program product, computer-readable storage medium, and electronic computing device

Abstract

The invention relates to a method for operating an industrial environment (10) comprising at least two different systems (12, 14) using an electronic computing device (16), having the steps of: - providing a machine learning algorithm (24) for semantically comparing a first parameter (20) of a first system (12) of the industrial environment (10) with at least one second parameter (22) of a second system (14) of the industrial environment (10); - providing a low-rank adaptation algorithm (26) for the machine learning algorithm (24); - semantically comparing the first parameter (20) with the second parameter (22); - adapting the low-rank adaptation algorithm (26) on the basis of the semantic comparison between the first parameter (20) and the second parameter (22); and - operating the industrial environment (10) on the basis of the machine learning algorithm (24) and the adapted low-rank adaptation algorithm (26). The invention also relates to a computer program product, to a computer-readable storage medium, and to an electronic computing device (16).

Description

[0001] Description

[0002] Method for operating an industrial environment with at least two different systems by means of an electronic computing device, computer program product, computer-readable storage medium and electronic computing device

[0003] The invention relates to a method for operating an industrial environment with at least two different systems by means of an electronic computing device according to the applicable patent claim 1. Furthermore, the invention relates to a computer program product, a computer-readable storage medium and an electronic computing device.

[0004] With an increasing number of interacting companies and a large number of suppliers for complex assemblies, the need for data interoperability is one of the greatest challenges for plant manufacturers, for example. The prerequisites for interoperability are the digital and actual representation of the system components using a uniform information model and ensuring the semantic interoperability of the systems. This describes a common understanding among the system components of the meaning of the exchanged information. An example of how semantic interoperability can be achieved is by implementing homogeneous semantics in the form of standardized vocabularies. Existing standards for defining terms and semantics, such as ECLASS, can be used for this purpose.However, compliance with these standards is not mandatory in the context of the Asset Administration Shell (AAS), which is an essential component of Digital Twins. Semantic diversity is represented in AAS information models through inheritance of so-called submodel elements.

[0005] Although standards are to be set, for example by ECLASS, it is to be expected that a heterogeneity of naming conventions will remain due to the large number of manufacturers, products and process participants. It is therefore necessary to be able to map different vocabularies to a standardized target semantics. In this context, mapping using Linked Data entails a high initial effort. This includes the time required to create a naming catalog, in particular the translation into different languages, as well as the maintenance of the catalog and the potential sources of error due to the manual transfer of information. Manually creating a mapping list at the attribute level is not proportionate, especially when there are changing suppliers that are to be integrated into the AAS system.

[0006] The object of the present invention is to provide a method, a computer program product, a computer-readable storage medium and an electronic computing device by means of which an automated mapping process for different parameters of different systems can be realized.

[0007] This object is achieved by a method, a computer program product, a computer-readable storage medium, and an electronic computing device according to the independent patent claims. Advantageous embodiments are specified in the subclaims.

[0008] One aspect of the invention relates to a method for operating an industrial environment with at least two different systems by means of an electronic computing device. A machine learning algorithm is provided for semantically comparing a first parameter of a first system in the industrial environment with at least one second parameter of the second system in the industrial environment by means of the electronic computing device. A low-level adaptation algorithm for the machine learning algorithm is provided by means of the electronic computing device. The first parameter is semantically compared with the second parameter by means of the machine learning algorithm. The low-level adaptation algorithm is adapted by means of the electronic computing device as a function of the semantic comparison of the first parameter with the second parameter.The industrial environment is operated depending on the machine learning model and the adapted low-level adaptation algorithm by means of the electronic computing device.

[0009] In particular, this makes it possible to implement automated mapping which reduces manual effort and can be provided with the help of models from the field of, for example, natural language processing. Within the scope of the present invention, for example, the use of Large Language Models (LLM) can be used. In this case, relevant information is based on various, predefined name combinations, for example manufacturer-specific, bsDD or the like. This not only enables the semantic comparison of the properties in an individual sub-model, but also provides input for the transformation of units of measurement, which refers to one of the main sub-problems. Overall, this eliminates the need for a name catalog.

[0010] In particular, the invention therefore provides that in order to determine the similarity of two different sub-model elements according to the Large Language Models, a model is required that can provide a similarity score. For example, the SBERT models can be used for this purpose. It refers to a so-called Siamese network that contains two or more identical sub-networks, where the models usually have exactly the same weights. Their goal is to calculate similarity values ​​of the inputs. Each BERT outputs pooled sentence embeddings. Pooling is a technique for generalizing features in a network, and in this case mean pooling works by averaging groups of features in the BERT model. After pooling, there are now two embeddings: one for sub-model element A and one for sub-model element B.When the model is trained, the model concatenates the two embeddings, which are then passed through a softmax classifier and trained with a softmax loss function. During inference – or when the model begins prediction – the two embeddings are compared using, for example, a cosine similarity function, which outputs a similarity score for the two sentences or words. In contrast to semantic matching with sentence transformers, when using large language models, an unknown sentence is fed along with a sentence from the target vocabulary to a model. This is done by integrating the names and descriptions of the sentences into a sentence structure, the prompt. This prompt is transformed into embeddings to make input processable for the large language model.

[0011] Each word of the input string is transformed into a numerical representation, specifically a vector, that describes the input label in a multidimensional feature space. These embeddings are then provided to the Large Language Model.

[0012] The low-rank adaptation algorithm is in particular a so-called LoRA model, in particular with so-called Low-Rank Adaptation of Large Language Models. In this case, it is particularly intended that, for example, matrices are changed while the actual machine learning algorithm remains unchanged. This means that base models can be provided as a machine learning algorithm and an adaptation of the low-rank adaptation algorithm can be carried out without, for example, influencing the weights of the machine learning algorithm. Particularly with very large machine learning algorithms, this has the advantage that a corresponding adaptation does not have to be carried out, since the low-rank adaptation algorithm is simply adapted "parallel" to the machine learning algorithm.In particular, in contrast to the state of the art, the solution developed here does not aim at a complete fine-tuning or relearning of the existing learning models, in particular the existing machine learning algorithm. Rather, small adjustments to the underlying neural network, in particular the low-rank adjustment algorithm, or the integration of existing methods such as Parameter Efficient Finetuning (PEFT) are provided as a lightweight parameterization of the results. Furthermore, the approach is not based on models with a small parameter set, but on the use of established multilingual Large Language Models. Another distinguishing feature is the continuous update cycle using Reinforcement Learning with Human Feedback (RLHF), which aims to continuously improve the solution proposed by the Large Language Model.

[0013] According to an advantageous embodiment, the machine learning algorithm is provided as a large language model. The large language model is in particular the Large Language Model. In other words, a large language model can be provided, whereby a simplified matching of the parameters can be carried out. Thus, on the basis of the vocabulary of the large language model, an improved matching or mapping of the parameters can be carried out. In order to carry out only fine-tuning, the low-level adaptation algorithm is used in parallel to the large language model. Thus, the large language model with a large vocabulary can be used and, at the same time, fine-tuning to the respective specific industrial environment can be carried out on the basis of the low-level adaptation algorithm.

[0014] A further advantageous embodiment provides that at least one weighting is adapted within the low-level adaptation algorithm. In particular, the low-level adaptation algorithm can be provided, for example, in the form of a neural network. Corresponding weights of the neural network can then be adjusted or adapted to perform the adaptation. Thus, an adaptation of the low-level adaptation algorithm can be realized in a simple manner.

[0015] It is also advantageous if the lower-rank adaptation algorithm is provided with at least two adaptable matrices. In particular, the at least two adaptable matrices can then be multiplied with each other to obtain a higher-rank matrix. The higher-rank matrix can then be used, in particular, for the corresponding weightings within the lower-rank adaptation model.

[0016] It is further advantageous if a low-level adaptation algorithm is provided with at least one encoder and / or one decoder. In particular, a parameter-efficient fine-tuning technique, which is classified as reparameterization, can thus be provided. For example, the input request can be converted into tokens, which are then converted into embedding vectors and forwarded to the encoder and / or decoder parts of the transformer. Thus, a reliable low-level adaptation of the large language model can be provided.

[0017] A further advantageous embodiment provides that the low-level adaptation algorithm is provided with at least one self-observation network and one feed-forward network. In particular, these two components can be used with two types of neural networks. In particular, self-observation networks and feed-forward networks can be used. The weights of these networks are learned during pre-training. After the embedding vectors of the inputs have been created, they are fed into the self-attention layers, where a series of weights is applied to calculate the attention values. During full fine-tuning, each parameter in these layers is updated.LoRA is a strategy that reduces the number of parameters to be trained during fine-tuning by freezing all of the original model parameters and then injecting a pair of boundary decomposition matrices next to the original weights. The dimensions of the smaller matrices are set so that their product is a matrix with the same dimensions as the weights they modify. These parameters must be adjusted depending on the specific large language model to be applied. The original weights of the large language model are then frozen, and the small matrices are trained using the supervised learning procedure.

[0018] A further advantageous embodiment provides that the machine learning algorithm is provided as a predefined machine learning algorithm. In this case, a predefined machine learning algorithm is to be understood in particular as a basic machine learning algorithm. In other words, prefabricated machine learning algorithms can be used in order to be able to carry out corresponding operation of the industrial environment. A specific adaptation of the predefined machine learning algorithm is not necessary. The specific implementation takes place in particular via the low-level adaptation algorithm. Thus, already established machine learning algorithms can be used, whereby improved operation of the industrial environment can be realized.

[0019] It is also advantageous if the low-level adaptation algorithm is also adapted into reinforcement learning with human feedback based on an optimization of results. This method is based in particular on the fact that the results delivered by the Large Language Model and LoRA can be further improved and the corresponding semantics can be mapped even more precisely. With this method, too, it should be noted that, for example, only the parameters of the LoRA concept are adapted and no retraining of the Large Language Model itself takes place. The practical application can, for example, be outlined in such a way that two sub-models of the AAS are merged. These each consist of a set of attributes with partially matching attribute names. The Large Language Model plus the LoRA model generates a probability matrix.Sometimes there are direct matches, such as in the case of the terms security or capacity. This then results in a necessary adjustment of the LoRA weights. The user resolves the conflicts and assigns the labels correctly. A process is then initiated in which a model with the old LoRA values ​​and a model with the new LoRA values ​​are compared in order to optimize them. This can provide reinforcement learning, so that over the course of operation, a more detailed and specific operation of the industrial environment can be realized.

[0020] It is also advantageous if the optimization is based on the mean square error of the human feedback. In particular, the mean square error (MSE) can be used as an error metric to measure the deviation from the user definition. The adjustments to the LoRA weights are based on the mean square error and are iteratively adjusted to approximate the ideal result. If the result of the existing model cannot be exceeded, the existing model remains in use; otherwise, it is overwritten by the weight update.

[0021] Furthermore, it has proven advantageous if the at least two parameters are compared using a cosine similarity function. In particular, a similarity value for the two sentences or words can be output based on the cosine similarity function. This allows a comparison to be realized in a simple yet reliable manner.

[0022] Furthermore, it has proven advantageous if the industrial environment is transferred into a digital twin model of the industrial environment, wherein the at least two systems are provided on the basis of the semantic comparison in the digital twin model. In particular, this can thus essentially provide a display for a user of the electronic computing device or the industrial environment. The systems can then be represented accordingly in the digital twin model. For example, corresponding parameters, in particular the recorded parameters, can be displayed accordingly. Based on the result of the electronic computing device, corresponding values ​​can be matched or mapped so that a user of the electronic computing device can easily see which parameters are currently being provided by the systems.For example, if a parameter value from a first system is a wattage and a power number in kilowatts is to be specified from a second system, this can now be displayed on the basis of the mapping in such a way that a user can reliably understand which actual power is being queried in a common display of the correct dimension within the systems.

[0023] It is also advantageous if at least the low-level adaptation algorithm is provided as a neural network. Here and in the following, an artificial neural network can be understood as software code that is stored on a computer-readable storage medium and represents one or more networked artificial neurons or can emulate their function. The software code can also contain several software code components, which can, for example, have different functions.In particular, an artificial neural network may implement a non-linear model or a non-linear algorithm that maps an input to an output, where the input is given by an input feature vector or an input sequence and the output may include, for example, an output category for a classification task, one or more predicted values, or a predicted sequence.

[0024] The method presented is, in particular, a computer-implemented method. Therefore, a further aspect of the invention relates to a computer program product with program code means that cause an electronic computing device, when the program code means are processed by the electronic computing device, to carry out a method according to the preceding aspect.

[0025] Furthermore, the invention therefore also relates to a computer-readable storage medium with at least the computer program product.

[0026] Yet another aspect of the invention relates to an electronic computing device for operating an industrial environment with at least two different systems, with at least one machine learning algorithm and a low-level adaptation algorithm, wherein the electronic computing device is designed to carry out a method according to the preceding aspect. In particular, the method is carried out by means of the electronic computing device.

[0027] Furthermore, the invention also relates to an industrial environment with at least two different systems and the electronic computing device according to the previous aspect.

[0028] Advantageous embodiments of the method are to be regarded as advantageous embodiments of the computer program product, the computer-readable storage medium, the electronic computing device, and the industrial environment. The electronic computing device and the industrial environment have material features for this purpose in order to be able to carry out corresponding method steps.

[0029] A computing unit / electronic computing device can be understood, in particular, as a data processing device containing a processing circuit. The computing unit can therefore, in particular, process data to perform computing operations. This may also include operations for performing indexed access to a data structure, for example, a look-up table (LUT).

[0030] The computing unit can in particular contain one or more computers, one or more microcontrollers and / or one or more integrated circuits, for example one or more application-specific integrated circuits (AS ICs), one or more field-programmable gate arrays (FPGAs), and / or one or more single-chip systems (SoCs). The computing unit can also contain one or more processors, for example one or more microprocessors, one or more central processing units (CPUs), one or more graphics processing units (GPUs) and / or one or more signal processors, in particular one or more digital signal processors (DSPs).The computing unit may also include a physical or virtual network of computers or other of the aforementioned units.

[0031] In various embodiments, the computing unit includes one or more hardware and / or software interfaces and / or one or more memory units.

[0032] A memory unit can be a volatile data memory, for example a dynamic random access memory (DRAM) or a static random access memory (SRAM), or a non-volatile data memory, for example a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a flash memory or flash EEPROM, a ferroelectric random access memory (FRAM), a magnetoresistive random access memory (MRM),MRAM (magnetoresistive random access memory) or phase-change random access memory, PCRAM (phase-change random access memory).

[0033] For use cases or application situations that may arise during the method and which are not explicitly described here, it may be provided that, in accordance with the method, an error message and / or a request to enter user feedback is issued and / or a default setting and / or a predetermined initial state is set.

[0034] Regardless of the grammatical gender of a particular term, persons with male, female or other gender identities are included.

[0035] Further features and combinations of features of the invention emerge from the figures and their description, as well as from the claims. In particular, further embodiments of the invention do not necessarily have to contain all features of one of the claims. Further embodiments of the invention may have features or combinations of features that are not mentioned in the claims. Herein:

[0036] FIG 1 is a schematic block diagram according to an embodiment of an industrial environment; and

[0037] FIG 2 is a schematic block diagram according to an embodiment of an electronic computing device.

[0038] The invention is explained in more detail below with reference to specific exemplary embodiments and associated schematic drawings. In the figures, identical or functionally equivalent elements may be provided with the same reference numerals. The description of identical or functionally equivalent elements may not necessarily be repeated for different figures.

[0039] FIG. 1 shows a schematic block diagram according to one embodiment of an industrial environment 10. The industrial environment 10 has a first system 12 and a second system 14 that is different from the first system 12. The industrial environment 10 has at least one electronic computing device 16. The electronic computing device 16 can, for example, have a display device 18. Results of the electronic computing device 16 can, for example, be displayed on the display device 18.

[0040] In particular, a first parameter 20 can be transmitted from the first system 12 to the electronic computing device 16. In particular, a second parameter 22 can be transmitted from the second system 14 to the electronic computing device 16.

[0041] FIG. 2 shows a schematic block diagram according to one embodiment of the electronic computing device 16. In the present exemplary embodiment, the electronic computing device 16 has at least one machine learning algorithm 24 and a low-level adaptation algorithm 26. Furthermore, the electronic computing device 16 can have an input layer 28 and an output layer 30.

[0042] According to one embodiment of the method, the machine learning algorithm 24 is provided for the semantical comparison of the first parameter 20 with the second parameter 22. For this purpose, the low-level adaptation algorithm 26 is also provided for the machine learning algorithm 24, in particular in parallel as shown here. The semantical comparison of the first parameter 20 with the second parameter 22 then takes place by means of the machine learning algorithm 24. The low-level adaptation algorithm 26 is then adapted in turn as a function of the semantic comparison of the first parameter 20 with the second parameter 22 by means of the electronic computing device 16, and the industrial environment 10 is operated as a function of the machine learning algorithm 24 and the adapted low-level adaptation algorithm 26.

[0043] In particular, it is provided that the machine learning algorithm 24 is provided as a large language model. Furthermore, it can be provided that at least one weighting is adapted within the low-order adaptation algorithm 26. The low-order adaptation algorithm 26 can be provided with at least two adaptable matrices. Furthermore, the low-order adaptation algorithm 26 can be provided with at least one encoder device 34 and one decoder device 36. In this case, as shown in FIG 2, the low-order adaptation algorithm 26 is preferably provided as a neural network. The low-order adaptation algorithm 26 can be provided with at least one self-observation network and one feedforward network.

[0044] The machine learning algorithm 24 can, for example, be

[0045] Large Language Model can be provided. Furthermore, it can be provided that the low-level adaptation algorithm 26 is additionally adapted based on an optimization of results in a reinforcement learning 32 with human feedback.

[0046] In particular, the figures show that, to determine the similarity of two different submodel elements (SE), for example, a model is needed that can provide a similarity score. One of the most commonly used examples is SBERT. It refers to a so-called Siamese network containing two (or more) identical subnetworks, where the models usually have exactly the same weights. Their goal is to calculate similarity values ​​of the inputs.

[0047] Each BERT outputs pooled sentence embeddings. Pooling is a technique for generalizing features in a network, and in this case, mean pooling works by averaging groups of features in the BERT model. After pooling, there are now two embeddings: one for submodel element A and one for submodel element B. When training the model, the model concatenates the two embeddings, which are then passed through a softmax classifier and trained with a softmax loss function. During inference—or when the model begins predicting—the two embeddings are then compared with a cosine similarity function, which outputs a similarity score for the two sentences or words. Unlike semantic matching with sentence transformers, when using LLMs (Large Language Models), an unknown SE is fed into the model along with an SE of the target vocabulary.To do this, the names and descriptions of the SEs are integrated into a sentence structure, the prompt. This prompt is then converted into embeddings to make the input processable for the LLM.

[0048] Each word of the input string is transformed into a numerical representation (vector) that describes the input string in a multidimensional feature space. These embeddings are then provided to the LLM.

[0049] LoRA is a parameter-efficient fine-tuning technique known as reparameterization. The input prompt is converted into tokens, which are then converted into embedding vectors and passed to the encoder and / or decoder parts of the transformer. These two components employ two types of neural networks: self-attention networks and feedforward networks. The weights of these networks are learned during pre-training. After the input embedding vectors are created, they are fed into the self-attention layers, where a set of weights is applied to calculate the attention scores. During full fine-tuning, every parameter in these layers is updated.LoRA is a strategy that reduces the number of parameters to be trained during fine-tuning by freezing all original model parameters and then injecting a pair of rank decomposition matrices alongside the original weights. The dimensions of the smaller matrices are set so that their product is a matrix with the same dimensions as the weights they modify. This parameter must be adjusted depending on the specific LLM to be applied. The original LLM weights are then frozen and the smaller matrices are trained using the supervised learning procedure.

[0050] For inference, the two low-rank matrices are multiplied together to create a matrix with the same dimensions as the frozen weights. These are then added to the original weights and replaced with these updated values ​​in the model. Now there is a fine-tuned LoRA model that can perform specific semantic matching. Since most of the parameters of LLMs are located in the attention layers, the greatest savings in trainable parameters can be achieved when LoRA is applied to these weight matrices. LoRA is applied for each matching of two semantics, to be adapted to the specific application task during the training process.

[0051] Typically, the use of an LLM begins with a computationally intensive training process of the LLM itself. Since this approach aims to reduce computational resources rather than training an LLM for each semantic matching, the use of already established LLMs is suggested. This could be a generalized LLM such as GPT or already specialized models for technical data.

[0052] During the training process, previously known semantic matches are used to initialize the values ​​of the matrices as part of the LoRA approach. This process uses matching semantic pairs as part of a probability matrix.

[0053] In this matrix, each value indicates the semantic similarity between the attributes in the rows and columns, where, for example, 1.00 represents a perfect match. The matrix is ​​the result of a pairwise query on the LLM when the two words or descriptions are similar, yielding a similarity value between zero and one.

[0054] The similarity score itself is based on cosine similarity, analogous to the SBERT approach. As a result, the LoRA weights are adjusted to approximate the real-world semantic matches as closely as possible.

[0055] Since incorrect semantic assignments, new semantics to be assigned, or other problems may still arise during productive use, the reinforcement learning method 32 with human feedback is applied to achieve further improvements. This method is based on further improving the result delivered by LLM + LoRA and mapping the corresponding semantics even more precisely. With this method, too, it should be noted that only the parameters of the LoRA concept are adjusted and no retraining of the LLM itself takes place. The practical use case can be outlined as follows: Two submodels of the AAS (Asset Administration Shell), i.e. the industrial environment 10, are to be merged. These each consist of a set of attributes with partially matching attribute names. The LLM+LoRA generates a probability matrix as shown below.Sometimes there are direct correspondences, as in the case of safety or capacity. The corresponding mapping of the terms would be as follows:

[0056] Ef fi ciency->Performance Safety->Safety Reliability->Reliability Capacity->Reliability already through 1 : 1 relationship / higher probability of another word

[0057] Cost->efficiency is already achieved by 1 : 1 relationship / higher probability for other word longevity->longevity

[0058] This requires adjusting the LoRA weights. The user resolves the conflicts and assigns the labels correctly. A process is then initiated that compares a model with the old LoRA values ​​and a model with the new LoRA values ​​to optimize them.

[0059] The mean squared error (MSE) is used as an error measure to measure the deviation from the user definition. The adjustments to the LoRA weights are iteratively adjusted to approach the ideal result. If the result of the existing model cannot be exceeded, the existing model remains in use; otherwise, it is overwritten by the weight updates. When matching between new semantics, the base model must be applied, which must then be adapted to the respective circumstances in the corresponding sub-method. The LoRA values ​​are only adjusted here after the first user adaptation, since the semantic matching of the LLM without LoRA can also be sufficient. The LoRA weights are therefore initialized so that they have no influence.

[0060] Reference symbol list

[0061] industrial environment first facility second facility electronic computing device

[0062] Display device first parameter second parameter machine learning algorithm low-level adaptation algorithm

[0063] Input layer

[0064] From there layer

[0065] Reinforcement learning

[0066] Encode cleaning

[0067] Decoder device

Claims

Patent claims 1. A method for operating an industrial environment (10) with at least two different systems (12, 14) by means of an electronic computing device (16), comprising the steps: - providing a machine learning algorithm (24) for semantically comparing a first parameter (20) of a first installation (12) of the industrial environment (10) with at least one second parameter (22) of a second installation (14) of the industrial environment (10) by means of the electronic computing device (16); - providing a low-level adaptation algorithm (26) for the machine learning algorithm (24) by means of the electronic computing device (16); - Semantically comparing the first parameter (20) with the second parameter (22) by means of the machine learning algorithm (24); - adapting the low-level adaptation algorithm (26) as a function of the semantic comparison of the first parameter (20) with the second parameter (22) by means of the electronic computing device (16); - Operating the industrial environment (10) in dependence on the machine learning algorithm (24) and the adapted low-level adaptation algorithm (26) by means of the electronic computing device (16).

2. The method according to claim 1, characterized in that the machine learning algorithm (24) is provided as a large language model.

3. Method according to claim 1 or 2, characterized in that at least one weighting is adapted within the low-order adaptation algorithm (26).

4. Method according to one of the preceding claims, characterized in that the low-rank adaptation algorithm (26) is provided with at least two adaptable matrices.

5. Method according to one of the preceding claims, characterized in that the low-level adaptation algorithm (26) is provided with at least one encoder device (34) and one decoder device (36).

6. Method according to one of the preceding claims, characterized in that the low-level adaptation algorithm (26) is provided with at least one self-observation network and one feedforward network z.

7. The method according to any one of the preceding claims, characterized in that the machine learning algorithm (24) is provided as a predetermined machine learning algorithm (24).

8. Method according to one of the preceding claims, characterized in that the low-level adaptation algorithm (26) is additionally adapted on the basis of an optimization of results in a reinforcement learning (32) with human feedback.

9. The method according to claim 8, characterized in that the optimization is carried out on the basis of a mean square error of the human feedback.

10. Method according to one of the preceding claims, characterized in that the at least two parameters (20, 22) are compared by means of a cosine similarity function.

11. Method according to one of the preceding claims, characterized in that the industrial environment (10) is converted into a digital twin model of the industrial environment (10), wherein the at least two systems (12, 14) are provided on the basis of the semantic comparison in the digital twin model.

12. Method according to one of the preceding claims, characterized in that at least the low-level adaptation algorithm (26) is provided as a neural network.

13. A computer program product comprising program code means which cause an electronic computing device (16) to carry out a method according to one of claims 1 to 12 when the program code means are processed by the electronic computing device (16).

14. A computer-readable storage medium comprising at least one computer program product according to claim 13.

15. Electronic computing device (16) for operating an industrial environment (10) with at least two different systems (12, 14), with at least one machine learning algorithm (24) and a low-level adaptation algorithm (26), wherein the electronic computing device (16) is designed to carry out a method according to one of claims 1 to 12.

Images