Computer-implemented method for using knowledge from a trained global large language model, computer program product and device

Abstract

Collecting task-specific knowledge for a classification task by AI in the industrial environment requires a high degree of expert knowledge about the respective area. For the purpose of formalisation and embedment in the machine learning model used, ML experts are however required. Close cooperation between data scientists and domain experts for each specific application is a prerequisite. This process is very costly and prone to errors in the event of incorrect communication, and as a result, these approaches are practical only for a limited number of applications where the benefit of solving the problem with an ML approach with a high degree of domain knowledge integration is sufficiently high to cover this cost. The proposed method overcomes the above-mentioned hurdles by providing a way of using knowledge contained in LLMs in small application-specific models. This is done by distilling expert knowledge, which is combined into rules for the direct classification and / or extraction of relevant data features in LLMs using domain-specific input requests and is then used in small task-specific models during inference or training.

Description

[0001] Computer-implemented method for using knowledge from a trained

[0002] Global Large Language Model, computer program product and device

[0003] A so-called Large Language Model (LLM) is a language model characterized by its ability to understand and generate general-purpose language. LLMs acquire these capabilities by learning statistical relationships from very large sets of text documents during a computationally intensive self-supervised and semi-supervised training process. LLMs are so-called neural networks that follow a transformer architecture.

[0004] As autoregressive language models, they work by taking an input text and repeatedly predicting the next token or word. Until 2020, a so-called "fine-tuning" process was the only way to adapt a model to perform specific tasks. Larger models can achieve similar results using so-called prompt engineering. They are assumed to acquire knowledge about syntax, semantics, and ontology contained in human language corpora, but also inaccuracies and biases present in those corpora.

[0005] Notable examples are OpenAI's GPT models (GPT, Generative Pretrained Transformer) (e.g., GPT-3.5 and GPT-4 used in ChatGPT), Google's PaLM (used in Bard), Meta's LLaMA, and Anthropic's BLOOM, Ernie 3.0 Titan, and Claude 2.

[0006] Recent developments in large language models are leading to their use not only in natural language analysis, but also in other areas of application. For example, there is considerable interest in exploring their potential use in industry.

[0007] However, unlike its application to language processing tasks for programming or natural languages, the utilization of its potential for discriminatory tasks with different data modalities, such as time series or tabular data, remains unclear. Examples of such applications include condition monitoring based on time series data, which is frequently used in industrial settings, process optimization based on tabular data, or various forms of quality assurance in the production of goods. Such data classification is still often performed using expert systems.

[0008] The following factors were identified as hindering the use of LLMs for these applications:

[0009] • The size of the models makes the use of these systems for certain tasks, e.g. on edge devices, difficult or unrealistic due to requirements on the required resources (storage space, computing capacity, etc.).

[0010] • Furthermore, training and adapting global LLMs for very specific tasks is inefficient and therefore comparing the model's conclusion with empirical observations is not possible in these cases.

[0011] • The great potential of current models can only be directly utilized for very specific data types, such as speech. The transfer of the results to other modalities, such as sensor data (in the form of time series, tables, image data, etc.), has so far been little investigated.

[0012] • The use of task-specific data, either for training or for inputting these models, may also raise privacy concerns, either on the side of the manufacturer or provider or on the side of the customers if the system is integrated into a product.

[0013] • The phenomenon of hallucination (alternatively called confabulation) is now known in the field of generative artificial intelligence (AI) as a black box system. This is a convincingly formulated result of AI that does not appear to be justified by training data and may be objectively false. Considering legal and non-functional requirements, this effect drastically complicates their use in autonomous systems in many applications, especially in industrial ones.

[0014] Since the release of very powerful LLMs, there have been initial activities to test these systems for a variety of use cases.

[0015] Example: LLM for “Weak Supervision”

[0016] In “Language Models in the Loop: Incorporating Prompting into Weak Supervision,” R. Smith, J.A. Fries, B. Hancock, and S.H. Bach describe the use of an LLM in combination with task-specific prompts to classify unlabeled input samples for a target application. The results serve as weak labels that are combined in a probabilistic model. This follows, for example, the paradigm of programmatic data labling introduced by A. Ratner, C. De Sa, S. Wu, D. Selsam, and C. Re, “Data Programming: Creating Large Training Sets, Quickly.” arXiv, Jan. 08, 2017, and used in products from “Snorkel” https: / / snorkel.ai / . These weak labels are then used to train a smaller target language model (LM).

[0017] In “PRBoost: Prompt-Based Rule Discovery and Boosting for Interactive Weakly-Supervised Learning.” arXiv, Mar. 18, 2022, R. Zhang, Y. Yu, P. Shetty, L. Song, and C. Zhang formulate rule templates instead of directly using prompts to generate weak labels for specific cases. The templates, in combination with specific data instances, are used to query an LM to generate labeling functions for text data that generalize beyond individual instances. The authors use a bootstrapping approach to find new rules for examples with high LM error.

[0018] Both approaches listed above feed data instances from the target application directly into the pre-trained LM. This results in the requirement that the input data for the target application must be text, which prevents the application of these approaches to the other use cases mentioned in the introduction, such as number sequences.

[0019] Incorporating knowledge about weak labels reduces the labeling effort for the target task, but training a model on a weak supervisory signal provides little or no benefit in terms of interpretability. Furthermore, incorporating weak labels, generated either by labeling functions or directly by the LM, could increase the data requirements of the target model, which can be very detrimental in industrial applications with low-data regimes.

[0020] Knowledge integration into training goal and inference

[0021] To make ML models more interpretable and potentially reduce the data requirements during training, expert knowledge can be directly incorporated into the preprocessing of the data, more specifically the feature extraction, or even the architecture or learning objective of the model.

[0022] In "Integrating Automated Knowledge Extraction with Large Language Models for Explainable Medical Decision-Making" by Haodi Zhang et al. (XP034518570), an LLM is combined with an MLN (Markov Logic Network). In a first step ("Knowledge Acquisition"), domain knowledge is collected and summarized using an LLM. In a second step ("Knowledge Formalization"), the collected knowledge is translated into first-order logic, creating an MLN. The third step involves inference with the MLN.

[0023] Furthermore, there are concrete approaches for incorporating domain knowledge into the learning objective, such as Physics-Informed ML, which uses differential equations in the learning objective, or Neuro-Symbolic ML, which uses task-specific logical rules. More generally, domain knowledge is introduced into the learning objective as an additional constraint, leading to a synchronization between the model's reasoning and the available domain knowledge, thereby increasing its knowledge conformity and thus the interpretability of the generated results.

[0024] MF Pradier, J. Zazo, S. Parbhoo, RH Perlis, M. Zazzi, and F. Doshi-Velez, in “Preferential Mixture-of-Experts: Interpretable Models that Rely on Human Expertise as much as Possible.” use human-based rules directly during the inference phase by incorporating them into the model architecture.

[0025] Gathering the task-specific knowledge required for these approaches often requires a high level of domain expertise. Formalizing this knowledge and embedding it in the machine learning model used, however, requires ML experts. The above points require close collaboration between data scientists and domain experts for each specific application. This process is very costly and prone to errors due to miscommunication, making these approaches practical only for a limited number of applications where the benefit of solving the task with an ML approach with a high degree of domain knowledge integration is sufficiently high to cover these costs.

[0026] The object of the invention is to provide a solution that enables the use of large language models even in specific applications in industrial environments. This should enable domain experts to use them even without in-depth knowledge of the functionality of the ML model.

[0027] This object is achieved by a computer-implemented method according to the features of patent claim 1.

[0028] Furthermore, this object is achieved by a computer program product according to claim 7 and a device according to claim 8. The object is achieved by a method for using knowledge from a trained global large language model for a target domain in industrial applications based on industrial processes and measured values ​​​​taken therefrom to solve prediction tasks with the following steps:

[0029] - Applying target domain-specific prompts to the Global Large Language Model, and

[0030] - Extracting target domain-specific knowledge from the global LLM in the form of rules for feature extraction from input data for a target domain-specific template model architecture,

[0031] - Integration of the extracted rules into a target domain-specific template model architecture,

[0032] - Training the target domain-specific template model architecture on application-specific data, and

[0033] - Create a small application-specific ML system for use in the target domain.

[0034] Advantageous embodiments of the invention are specified in the subclaims.

[0035] In addition to the extracted rules for feature extraction, in one embodiment of the invention, classification rules can be generated for input data for the application-specific ML system by the Global Large Language Model, for performing a classification task in the domain-specific industrial

[0036] Application.

[0037] These classification rules ensure that the result of the task to be solved by the system, for example the prediction task in the industrial plant, will be more accurate.

[0038] A prediction task can be imagined, for example, as the determination of wear in machining tools.

[0039] The classification rules generated in this way are then integrated into the application-specific ML system and trained into a target model.

[0040] In a further embodiment, the integration of generated

[0041] Classification rules into the application-specific ML system using a lead model, where the lead model performs an assignment of input data to an ML-based classification model or one of the classification rules.

[0042] This design alternatively decides whether an input is processed by the pre-trained ML-based classification model or by a rule. Rules have the advantage that they require less computing power and, unlike the ML-based classification model, do not represent a "black box," thus the result is transparent.

[0043] Feedback from a domain-specific trained ML system provides information about the application of the rule and the correctness of the results when applying the generated rules for feature extraction or direct classification to the input data. This information can be used to refine the prompts, which are then required to generate rules for feature extraction or direct classification.

[0044] The feedback also ultimately ensures an improvement in the final result or the correctness of the results for the prediction task.

[0045] If the generated program code for the rules is immediately put into a predefined form (i.e. “formalized”), then it is advantageous if it has at least one interface for direct integration into the modular, application-specific ML system.

[0046] This should be possible by formulating the correct prompt, which then saves additional integration and testing effort.

[0047] The classification task of the method can be the condition monitoring of a device in the industrial plant (400) based on measured values ​​taken from the device, for example, by a sensor. Alternatively, it would also be possible to record and evaluate all possible operating data through the device's control system. This approach, also called predictive maintenance, leads to overall lower maintenance costs and higher availability in the plant, since maintenance is not performed too early.

[0048] Examples of applications could be (the list is not exhaustive):

[0049] - Detect faults

[0050] - Locate disturbances

[0051] - Identify causes of faults,

[0052] Detect throughput fluctuations and schedule maintenance and repairs.

[0053] Another embodiment offers a domain-specific prompt library, which offers a set of domain-specific prompts to choose from. This can be implemented, for example, based on an application description.

[0054] The proposed method overcomes the above-mentioned hurdles by providing a way to utilize knowledge contained in LLMs in small, application-specific models. This is achieved by distilling the expert knowledge contained in LLMs into rules for direct classification and / or extraction of relevant data features using domain-specific prompts, and subsequently using it during inference or training in small, task-specific models.

[0055] In summary, the interpretability of task-specific models is increased and the data and label requirements are reduced by extracting the knowledge valuable for solving the task from a pre-trained, very general LLM.

[0056] The potential of LLMs can also be utilized in small models, which have the advantage of significantly reduced requirements in terms of the required "computing power" (computing capacity, storage space) and can therefore also be used on limited computing resources (such as the edge devices commonly used in industrial environments).

[0057] The described solution, which incorporates universal LLMs into this process, reduces the time required for domain experts and data scientists to integrate knowledge, thus reducing the effort required to use machine learning approaches for the described applications.

[0058] A system is proposed that incorporates knowledge extracted from LLMs into a smaller target model, independent of the data modality of the task.

[0059] Preferred embodiments of the invention are illustrated below by the figures.

[0060] Figure 1 a high-level system architecture

[0061] Figure 2 shows an exemplary implementation for an industrial plant.

[0062] The procedure shown in Figure 1 describes the information flow under consideration. This is essentially divided into three sections,

[0063] - an application-specific part, 30

[0064] - a domain- or task-specific part, 20 and

[0065] - a domain-independent part 10 (plus 20 if necessary)

[0066] The following distinction is used below:

[0067] Task- or domain-specific 20 are the steps that generally remain the same for a particular task to be solved, e.g., extracting rules for the detection of bearing damage based on vibration data.

[0068] Application-specific, 30, is the training of the model on data that comes from a specific plant / machine / asset and thus the application-specific model that is optimized for this specific application.

[0069] A universal Large Language Model (LLM) is used, which is trained with a very large dataset 101, 104 comprising data from a variety of different domains 102, and hopefully also from the desired domain 103, for example, (scientific) publications, for the classification task of condition monitoring of a specific component, such as bearings. This can be, for example, a general-purpose model such as ChatGPT or an LLM optimized for industrial applications. However, it is assumed that the LLM has also gained insights into the target domain within its cross-domain knowledge during training with the entire dataset.The trained domain-specific knowledge can be based on or draw on domain-specific literature such as scientific publications, discussions in a specialist forum, manuals or other literature related to the target or application area.

[0070] To extract the target-domain-specific knowledge from the LLM, a second, domain-specific part of the procedure uses target- or application-domain-specific prompts 201 (so-called "prompts," i.e., short texts or instructions that can be entered to trigger specific responses or actions). The prompts are formulated as tasks to query the required target-domain-specific knowledge in the form of rules 202.

[0071] The prompts can also be formalized to solve prediction tasks (more specifically: classification tasks) based on industrial data (e.g., sensor data). Due to the code generation capabilities of the LLM, the prompts are formalized in such a way that the resulting target domain rules 203 directly lead to software functions, which can be integrated through a predefined, uniform interface directly into a target domain-specific template model architecture 205, defined modularly as a software module (step 204). The model template offers the possibility of integrating the extracted knowledge into the training or inference of the model. 300 represents a possible implementation of the integration.

[0072] In the application-specific section 30, the model is then trained using application-specific labeled data. The knowledge gained from the LLM is linked to empirical observations describing the task using application-specific training data. These observations could be, for example, collected vibration measurements describing the good and bad condition of a machine being monitored. By training the model template, which contains the extracted LLM knowledge, these two sources of information are linked and implicitly compared.

[0073] The knowledge integration method must account for noise, e.g., caused by hallucinations of the LLM, and the incompleteness of the extracted knowledge. This is the case when integrating the rules as constraints as an additional error term into the model's learning objective, or when directly integrating the knowledge as classification rules into the model's architecture.

[0074] An example of constraints that can be integrated into ML models are physical rules. For example, in a closed system, the predicted pressure should increase as the input variable, temperature, increases. Such simple relationships between output and input variables can be imposed on the model as additional constraints during training to fulfill the objective.

[0075] After the training step 302, the resulting small application-specific ML model can be deployed in the narrow target application, 304 e.g., on an edge device in an industrial plant.

[0076] The hierarchical structure depicted in Figure 1 illustrates the reusability of the individual components. While the global LLM, which only needs to be trained once in step 10 and can then be used in different application domains (step 20), the created template models are specified by the prompts used for a specific application type. This allows the template models to be transferred to different machines or environmental conditions, for example. After the training step with application-specific / machine-specific data, the model is adapted to a specific application. For example, a predetermined engine model containing a bearing is to be monitored. The classification of bearing failures of the bearings installed in this specific engine model (as mentioned in the example) is then an application- or target-specific training objective.

[0077] In the embodiment shown in Figure 1, a feedback loop is introduced between the generated model 303 and the query of the LLM 202. The feedback can include, for example, information regarding the quality of the generated rule with respect to the training data or additional prototypical data points where the rule applies / is used or does not apply / is not used. This enables the rules to be revised based on empirical observations / application-specific training data. By providing feedback from a larger number of different applications, a general applicability of the adapted rules can be ensured and overadaptation to a specific application can be prevented.

[0078] Figure 2 shows a possible embodiment of the proposed invention, in which the knowledge extracted from the LLM is integrated into a target model as classification rules that are evaluated during inference (i.e., runtime). To create the application model, in this embodiment, the LLM is deployed in the cloud 210 and provides an interface for querying outputs. Furthermore, there is an application-specific prompt library 211 consisting of prompts 212 to provide rules for solving the target task. Different prompts focus on generating rules based on different attributes of the input data.

[0079] The prompts contained in the prompt library 211 can originate from previous similar applications or be manually generated by domain experts for the current task. The prompts, in combination with the LLM, result in classification rules 214, which are provided in parallel with a small ML-based classification model 311. An additional, ML-based guidance model 312 assigns data instances either to one of the rules 214 or to the ML-based classification model 311. Using a set of labeled training data, the entire system is trained to optimize the target classification task. The guidance, in combination with the ML-based classification model, accounts for the incompleteness and noisiness of the extracted rules. On the other hand, the rules contained in the inference generalize beyond the training sets and can thus reduce the data requirements to obtain a robust model.

[0080] The application-specific trained system 902, consisting of the rules, the ML-based classification model 311, and the ML-based control model 312, can then be deployed directly ("on-premise") 400 in the industrial plant to perform the task, for example, the condition monitoring of an industrial process 900 using a sensor 901, directly on the factory floor. For instances covered by the rules, the described architecture increases the interpretability of the results by representing the corresponding classification rule.

[0081] A concrete example of the industrial process monitored in Figure 2 could be a running motor, such as one used in a machine tool or conveyor belt, with a vibration sensor mounted on the motor housing. The target task is to classify the condition of the bearings installed in the motor.

[0082] Bearing fault detection based on vibration data is discussed in both scientific literature and industrial publications, e.g., in data sheets and white papers from bearing manufacturers, or even in various scientific publications. Corresponding data and information are therefore very likely also trained in a global LLM. For this reason, an LLM trained on a vast excerpt of documents available on the internet also encodes domain-specific knowledge about bearing fault detection. The most commonly considered rules for classifying bearing faults are based on features that represent characteristic frequencies within the spectrum of the vibration signal. These frequencies, in turn, can be determined by the geometry of the bearing used.

[0083] In a further embodiment, only the features extracted via rules generated by LLMs could be used. In the above-described embodiment for checking for damage to a bearing, the extraction of the features occurs implicitly within the classification rules, e.g., as described via the geometry of the bearing used. A concrete example of a rule generated by an LLM for the above-described application of monitoring an engine bearing could look like this:

[0084] Prompt:

[0085] “Write a binary rule for classifying bearing failures in the envelope spectrum based on the provided failure frequency.

[0086] Output: 1 = damaged, 0 = not damaged;

[0087] Interface: fault_classification_1(vibrational_signal, sampling_frequency, fault_frequency, rotational_speed);

[0088] Set a reasonable threshold based on the signal characteristics.”

[0089] The following is an example program code for the classification of bearing faults: def fault_classif ication_l (vibrational_signal, sampling_f requency, f ault_f requency, rotational_speed) :

[0090] #compute the envelope of the signal using the Hilbert transform analytic_signal = Hilbert (vibrational_signal) envelope = np.abs (analytic_signal)

[0091] #compute the spectral content of the envelope using a Fast Fourier Transform (FFT) n_samples = len (vibrational_signal) frequencies = np . f ft . rf ftf req (n_samples, dl / sampling_frequency) envelope_spectrum = np. fft.rfft (envelope)

[0092] #find the index of the frequency closest to the fault frequency idx_fault_f requency = np.argmin (np.abs (frequencies - fault_f requency) )

[0093] #find the index of the frequency closest to the lx rotational speed frequency f req_rota [ tional_speed = rotational_speed / 60 idx_rotational_speed = np.argmin (np.abs (frequencies - f req_rotational_speed) ) #compute the ration between the spectral content at the fault frequency and the spectral content at the lx rotational speed frequency ratio = envelope_spectrum [ idx_f ault_f requency ] / envelope_spectrum

[0094] [ idx_rotational_speed]

[0095] #set a threshold for the ration based on s ignal characteristics ( e . g . s ignal to noise ratio, amplitude ) threshold = 0 . 5

[0096] #return 1 i f the ratio is above the threshold, otherwise return 0 i f ratio > threshold : return 1 else : return 0

[0097] Diese Regeln entsprechen im Prinzip denen in einem Expertensystem.

[0098] If this rule is selected by the control model 312 and it outputs a "1" and thus an indication of a damaged bearing, the expert knowledge encoded in the rule, which in this case is given by the choice of the amplitude of the characteristic frequency in the frequency domain, is considered applicable and the procedure takes the corresponding stored action: for example, sending a warning that leads to maintenance or even shutdown of the plant, or updating a system status database.

[0099] The insights derived from the training data account for the cases where the generated rules 213 are not applicable by relying directly on the ML-based classification model 311 when the rules are integrated into the model architecture, as shown in Figure 2, or through the loss when the rules are integrated into the learning objective of a dedicated ML model.

[0100] Due to the good code generation capabilities of current LLMs (and the expected further improvements in future versions, including the addition of further knowledge through training), a common interface for the rules in the prompts can be defined, thus allowing the user to incorporate new rules without any specific ML programming knowledge.

[0101] With further progress in the field of LLMs and their targeted adaptation to industrial use cases, it can be assumed that increasingly complex and reliable rules can be generated and applied in the future.

[0102] The procedure described above makes it possible to exploit the high potential of LLMs in applications where they are not currently used, especially in industrial applications with their increased requirements.

[0103] To our knowledge, there are currently no systems that attempt to apply the knowledge contained in LLMs to discriminatory tasks on data modalities beyond speech and images in small, application-specific models. This holds enormous potential for a wide range of use cases, for example, in the industrial environment described above. Distilling this knowledge into small, application-specific ML systems makes it possible to leverage an existing, well-established software framework for deployment, such as an industrial edge infrastructure. The ability to reconcile the rules derived from the LLM with empirical observations, thus explicitly accounting for noise and incompleteness, offers a way to deal with the tendency of LLMs to hallucinate false facts with high confidence.Considering this, one of the main obstacles to the use of LLMs in the industrial sector could be addressed.

[0104] List of reference symbols

[0105] LLM Global Large Language Model

[0106] 10 Target domain-independent step

[0107] 101 Task-independent set of text records

[0108] 102 Cross-domain knowledge

[0109] 103 Target domain-related knowledge

[0110] 104 Training

[0111] 20 Domain-specific step.

[0112] 200 feedbacks

[0113] 201 Task-specific prompt

[0114] 202 query rules

[0115] 203 Target Domain Rules for Feature Extraction and Classification

[0116] 204 Integration step

[0117] 205 ML System

[0118] 210 LLM in Cloud

[0119] 211 Task-specific prompt library

[0120] 212 Prompt

[0121] 213 Extracted task-specific rule library

[0122] 214 Extracted task-specific rule

[0123] 30 Application-specific step

[0124] 300 Model Training

[0125] 301 Application-specific data

[0126] 302 target domain-specific model (architecture)

[0127] 303 application-specific model system

[0128] 304 Provision

[0129] 305 Provided solution model

[0130] 310 training data

[0131] 311 Application-specific ML system

[0132] 312 Leading model

[0133] 400 Provision of the model

[0134] 900 industrial process

[0135] 901 Sensor

[0136] 902 Application-specific model on site

Claims

Patent claims 1. Computer-implemented method (20) for using knowledge from a trained Global Large Language Model (LLM, 210) for a target domain in industrial applications based on industrial processes and measured values ​​taken therefrom to solve prediction tasks, comprising the following steps: - Applying application domain-specific prompts (201 , 212) to the Global Large Language Model (LLM, 210), and Extracting target domain-specific knowledge (202) from the global LLM in the form of feature extraction rules (203) from input data for a target domain-specific template model architecture (205) and - Generation of classification rules (213, 214) for input data for the application-specific M-System by the Global Large Language Model (LLM, 210), to perform a classification task in the domain-specific industrial application Integration of the extracted rules and classification rules (213, 214) into a target domain-specific template model architecture (205), - T raining (302) of the target domain-specific template model architecture (204) on application-specific data, taking into account the classification rules (213, 214) and Creating a small application-specific ML system (303) for use (902) in the target domain (400), characterized in that the integration of classification rules (213, 214) into the application-specific ML system (303) is carried out using a lead model (312), wherein the lead model performs an assignment of input data to the ML-based classification model (311) or a classification rule (214).

2. Computer-implemented method according to claim 1, characterized in that feedback (200) is returned to a domain-specific trained ML system (303) with information about the application of the rule and the correctness of the results when applying the generated rules for feature extraction to the input data and is used to refine the prompts (211) for generating rules for extracting features or direct classification (214).

3. Computer-implemented method according to one of the preceding claims, characterized in that the Global Large Language Model (LLM, 210) is used to directly generate program code for classification rules and feature extraction, and the program code is formalized so that it has an interface for integration into the modular application-specific ML system.

4. Computer-implemented method according to one of the preceding claims, characterized in that the classification task comprises a condition monitoring (901) of at least one device (900) in the industrial plant (400) on the basis of measured values ​​taken from the device.

5. Computer-implemented method according to one of the preceding claims, characterized in that there is a domain-specific prompt library (211) comprising a set of prompts for selection based on an application description.

6. Computer-implemented method according to one of the preceding claims, characterized in that additional documents which were not present in the training (104) are provided to the Global Large Language Model (LLM, 210) during the query (202) in order to expand the domain-specific knowledge (103).

7. Computer program product suitable and arranged to carry out the steps of a method according to one of the preceding claims.

8. Device for using knowledge from a trained Global Large Language Model (LLM, 210) for a target domain in the industrial application based on industrial processes and measured values ​​extracted therefrom to solve prediction tasks with a small application-specific ML system (303) for use (902) in the target domain (400), which is generated by an ML model (204), for integrating extracted rules and generated classification rules for input data for the application-specific ML system (213, 214) by the Global Large Language Model (LLM, 210), for carrying out a classification task in the domain-specific industrial application, wherein the Classification rules (213, 214) are integrated into the target domain-specific template model architecture (205) and trained into a target model (303 300), by applying application domain-specific prompts (201 , 212) to the Global Large Language Model (LLM, 210), extracted target domain-specific knowledge (202) in the form of rules for feature extraction (203) from input data for an application-specific ML model (303) Training (302) of the target domain-specific template model architecture (205) on application-specific data, characterized in that the integration of classification rules (213, 214) into the application-specific ML system (303) is carried out with the aid of a lead model (312), wherein the lead model performs an assignment of input data to the ML-based classification model or a classification rule.

9. Device according to claim 8, characterized in that an improvement of the prompts (211) for generating rules for the extraction of features or direct classification (214) by returning information about the application of the rule and the correctness of the results when applying the generated rules for feature extraction by means of a feedback (200) and the domain-specific trained ML system (303) to the input data.

10. Device according to one of the preceding claims 8 or 9, characterized in that the Global Large Language Model (LLM, 210) is used to directly generate program code for classification rules and for feature extraction, and the program code is formalized so that it has an interface for integration into the modular application-specific ML system.

11. Device according to one of the preceding claims 8 to 10, characterized in that the classification task comprises a condition monitoring (901) of at least one device (900) in the industrial plant (400), based on measured values ​​taken from the device.

12. Device according to one of the preceding claims 8 to 11, characterized in that there is a domain-specific prompt library (211) comprising a set of prompts for selection based on an application description.

13. Device according to one of the preceding claims 8 to 12, characterized in that additional documents that were not present in the training (104) are provided to the Global Large Language Model (LLM, 210) during the query (202) in order to expand the domain-specific knowledge (103).

Images