Servicing a wind turbine using a computer implemented service instruction generator applying a trained natural language model

Abstract

The present disclosure pertains to service of a wind turbine based on a determined service instruction. The service instruction and a resulting service action is determined using a computer implemented service instruction generator applying a trained natural language model. The service instruction generator accesses a repository of service instructions and a repository of fault entries. At a defined recurring time a current text corpus is prepared based on the accessed service instructions and fault entries. The natural language model is trained on the current text corpus. The service instruction and a resulting service action is determined querying the natural language model trained on the current text corpus.

Description

[0001] SERVICING A WIND TURBINE USING A COMPUTER IMPLEMENTED SERVICE

[0002] INSTRUCTION GENERATOR APPLYING A TRAINED NATURAL LANGUAGE

[0003] MODEL

[0004] FIELD OF THE INVENTION

[0005] The present disclosure pertains to the field of servicing a wind turbine, and in particular it relates to performing a service action based on a determined service instruction.

[0006] BACKGROUND OF THE INVENTION

[0007] During service, maintenance and / or repair of a wind turbine, it is desirable to be able to reliably identify a correct action for improving the state of the wind turbine. Current methods include relying on experience of the service technician and / or written service manuals, wind turbine technical documents, service schedules, work instructions, records maintained in a log stored in a database, etc. A wind turbine is a very complex piece of equipment and identifying the correct service task may be very time consuming and involve a certain risk of either basing the service task on wrong information or not performing the best corrective action.

[0008] A wind turbine incorporates diagnostic controls and sensors that report faults when anomalous operating conditions of the equipment arise. Typically, to diagnose the problem, a technician will study the fault log to identify the nature of the problem and determine whether a repair is necessary, however for a complex piece of equipment such as a wind turbine, the amount of information in logs may be quite overwhelming.

[0009] Another service issue may be the availability of a service history or difficulties in extracting relevant information from a service history which may be documented in various systems that may be difficult to access on site. Moreover, while technical documentation may be shared for a group of wind turbines, a service history is typically directed to a single wind turbine only, nevertheless service actions performed for a given wind turbine may be relevant for another wind turbine under service.

[0010] It is against this background that the invention has been devised. SUMMARY OF THE INVENTION

[0011] In connection with servicing a wind turbine, it would be advantageous to be able to base the service action on as much information as possible in a time efficient manner. The inventor of the present invention has realized that a properly trained natural language model is efficient in extracting relevant information from a large corpus of material and may if properly arranged be used in an effective manner for providing service instructions relevant for perform a service action on a wind turbine.

[0012] Accordingly, in a first aspect, there is provided a method of servicing a wind turbine using a computer implemented service instruction generator applying a trained natural language model; the method comprises: query the service instruction generator; wherein service instruction generator performs: accessing a repository of service instructions of specific wind turbine models, each service instructions in the repository of service instructions having a service instruction status tag; accessing a repository of fault entries listing specific identified faults of specified wind turbines and associated repair actions done to mitigate the identified faults, each fault entry having a fault entry status tag; at a defined recurring time prepare a current text corpus, wherein the current text corpus comprises a plurality of corpus elements, and wherein the plurality of corpus elements comprises the service instructions in the repository of service instructions having a selected service instruction status tag and the fault entries in the repository of fault entries having a selected fault entry status tag; train the natural language model on the current text corpus; generate a service instruction using the natural language model based on the query; and perform a service action on the wind turbine based on the service instruction.

[0013] A wind turbine is a complex piece of equipment which needs service and maintenance when used. A properly serviced wind turbine is expected to operate for more than 20 years. As the wind turbine operates, a great number of modifications to the various components is expected and since the wind turbine operates in an ever-changing environment due to variability in wind, temperature and other environmental conditions, a service history of a given wind turbine is typically fairly unique. Service instructions for a wind turbine may therefore fast become obsolete if not continuously updated to match the actual state of the wind turbine.

[0014] A trained natural language model is a powerful tool which is capable of extracting relevant information from a large text corpus, however there is a risk that if not properly trained, a service instruction using the natural language model risk being based on only partly valid material.

[0015] The present invention provides the advantage that a service instruct! on / service action is generated from a trained natural language model which is trained on a current text corpus. The current text corpus is the updated text corpus at the defined recurring time comprising corpus elements which are valid for the wind turbine operation at the defined recurring time. In this manner it can be ensured that the service instruct! on / service action takes into account the specific updated state of the wind turbine.

[0016] While the current text corpus can be build based on various information, by including as a minimum at least service instructions and fault entries and associated repair actions, the basic information for performing an appropriate service action is present, and by providing a service instruction status tag to each service instructions in the repository and a fault entry status tag to each fault entry, it can be ensured that only relevant corpus elements are used in the current text corpus, in the form of corpus elements having a selected service instruction status tag and a selected fault entry status tag. In this manner, it can be detailed controlled which material that is used for the training and thereby ensure that only up-to-date and relevant material is used in the training of the natural language model.

[0017] A tag (such as the selected service instruction status tag and the selected fault entry status tag) can be seen as a structural information assigned to a data element. For example, a tag can be used for classification, categorization, and / or context.

[0018] The disclosed techniques advantageously provide an efficient training of the natural language model, e.g. by preparing a current text corpus including, inter alia, respective selected service instruction status tag and respective selected fault entry status tags. For example, the efficiency may be particularly appreciated when training is performed periodically, such as between hourly and weekly. The disclosed techniques advantageously provide an improvement in the accuracy of the service instructions provided as output by the natural language model. In other words, generating service instructions using a model trained on tagged data allows to improved granularity (e.g. per component tracking), clearer learning pathways to identify patterns and relationships, and higher accuracy in the generation of the service instruction. Further, the disclosed tagging allows for an increased robustness in that tagging reduces the likelihood of false positives / false negatives.

[0019] The status tag may define information relevant to the instruction or entry, such as if the instruction or entry is valid or obsolete, or the domain of validity. But the status tag may also define a quality of an instruction or entry, i.e. is the instruction or entry from a verified source, a skill level of the source, a trust level of the source, etc. Use of status tags may ensure that only the most relevant, high-quality, and appropriate data is included in the training corpus. This will help improve the performance and accuracy of the language model output.

[0020] In embodiments, the current text corpus further comprises corpus elements being sensor reading data from one or more sensors of the specific wind turbines. The sensor readings may be reading obtained in a specified period prior to preparing the current text corpus.

[0021] It is advantageous to include sensor readings in the text corpus. However, as a large number of sensor readings are present in a wind turbine, it may be advantageous to limit the sensor readings to a specific period. A properly trained natural language model is capable of detecting outlier data, which may be a valuable input into a service task at hand. The specific period may be the period since the last service of the wind turbine, the last few months, the last year, or any other appropriate period.

[0022] In embodiments, the current text corpus further comprises corpus elements listing components of the specific wind turbines. Optionally together with associated information related to at least the last service action performed on the components.

[0023] It is advantageous to also include a list of specific components of a given wind turbine thereby enabling a very specific service instruct! on / service action down to the component level. Moreover, by including associated information related to at least the last service action performed on the components the service instruct on / service action may take into account service actions performed at same component types of other wind turbines, as well as provide pre-emptive service instruction to deal with components which have been experienced faults at an earlier stage at other turbines or which exhibit a large likelihood of experiencing a fault in a near future.

[0024] In embodiments, the current text corpus further comprises corpus elements of ad hoc selectable text of the specific wind turbines.

[0025] By including ad hoc selectable text into the current text corpus relevant material which may be included by the service technician or other technical personnel, thereby ensuring that the service instruct! on / service action may be based on any type of material deemed important for the operation of the components of the wind turbine. The ad hoc selectable text may e.g. be provided as selected emails. The ad hoc selectable text may be manually selectable text by the service technician or other technical personnel. Such selectable text may be provided together with a status tag identifying the information as manually selectable text provided due a specific selection by a service technician or other technical personnel.

[0026] In an embodiment, the service instruction generator comprises a defined application programming interface (API) for receiving corpus elements to be included in the current text corpus.

[0027] Providing a defined API for receiving corpus element is an efficient manner of enabling both an automated setup for receiving corpus elements and a setup for receiving manual / ad hoc corpus elements. With a defined API, the service instruction generator may be provided with a versatile interface for receiving corpus elements from various and diverse information sources.

[0028] In embodiments, the text corpus is managed to only include unique corpus elements thereby controlling the data size of the text corpus and reduce the risk of putting emphasis on a given element via a multiple presence in the text corpus. This may be implemented by the service instruction generator to prepare the current text corpus by comparing a received corpus element with the corpus elements of the existing text corpus and include the received corpus element into the current text corpus if the received corpus element is not included in the existing corpus material.

[0029] In embodiments, the text corpus is dynamically maintained by only including corpus elements with a given status tag, hereunder to remove corpus elements where the status tag is changed to a status tag not to be included in the current text corpus. This may be implemented by the service instruction generator to prepare the current text corpus by identifying the status tag of the corpus elements of the existing text corpus and remove corpus elements from the current text corpus if the status tag of a given corpus element is different from a selected status tag.

[0030] In this manner is the content of the text corpus can be detailed controlled and it can be ensured that obsolete material is removed from the current text corpus in an efficient manner.

[0031] In embodiments, the defined recurring time is a predefined recurring time interval between an hour and a week. By setting the defined recurring time as a predefined time interval, it can be ensured that the text corpus is always up-to-date. In an advantageous embodiment the natural language model may be re-trained e.g. once a day.

[0032] If a wind turbine is operated in a very stable manner, the text corpus may change only slowly, and therefore there may not be a need for frequent re-training. In embodiments, the defined recurring time is determined based on an amount of received corpus element(s) to be included in the current text corpus.

[0033] In embodiments the service instruction further comprises a listing of components to be used for performing the service action. The embodiments of the present inventions may be implemented in a manner where the service technician has access to the service instruction generator in preparation for the service visit. By including a listing of components to be used for performing the service action, the service technician can ensure that the relevant tools and replacement components are available for the service action. In this manner the service technician does not risk travelling to the oftentimes far locations of the wind turbine site, just to realize that a relevant tool or component is missing. According to another aspect of the invention there is provided a computer program product comprising instructions which, when the program is executed by a computer, cause the computer to carry out the method according to the first aspect of the invention. The computer program may be provided embodied in a non-transient, computer-readable storage medium storing instructions thereon that when executed by a processor cause the processor to execute a method as defined above.

[0034] The computer program product may be provided on a computer readable storage medium or being downloadable from a communication network. The computer program product comprises instructions to cause a data processing system to carry out the instruction when loaded onto the data processing system.

[0035] According to another aspect of the invention there is provided a system for service of a wind turbine using a computer implemented service instruction generator applying a trained natural language model; the system comprises: a client device arranged for generating a query to be provided to the service instruction generator, and for receiving a generated service instruction from the service instruction generator; a processor; and a memory storing thereon executable instructions that, when executed, cause the processor to implement the service instruction generator in accordance with the first aspect of the invention.

[0036] In general, the system may be a unit or collection of functional units which comprises one or more processors, input / output interface(s) and a memory capable of storing instructions that can be executed by a processor.

[0037] In general the various aspects of the invention may be combined and coupled in any way possible within the scope of the invention. These and other aspects, features and / or advantages of the invention will be apparent from and elucidated with reference to the embodiments described hereinafter.

[0038] BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Embodiments of the invention will be described, by way of example only, with reference to the drawings, in which Fig. 1 illustrates a flowchart representing general steps in accordance with embodiments of the present invention;

[0040] Fig. 2 illustrates a flowchart representing general steps performed by the service instruction generator in accordance with embodiments of the present invention;

[0041] Fig. 3 illustrates an example system for obtaining different types of corpus elements together with elements of a general wind turbine system; and

[0042] Fig. 4 illustrates a flow diagram of an example text corpus manager.

[0043] DESCRIPTION OF EMBODIMENTS

[0044] Figure 1 illustrates a flowchart representing general steps involved in the method of servicing a wind turbine using a computer implemented service instruction generator in accordance with embodiments of the present invention.

[0045] A query Q is sent 1 to the service instruction generator SIG. The query may be a text-based query or a voice-based query. As the natural language model is text-based, a voice-based query will be transferred into a text-based query using a voice recognition technique.

[0046] The service instruction generator is a computer implemented algorithm which manages the current text corpus CTC and handles the natural language model NLM, hereunder the training of the NLM. The service instruction generator is typically implemented as a collection of interconnected computing modules.

[0047] The output of the service instruction generator is a service instruction SI which is used for performing a service action SA. The service action may be a maintenance task, a repair task or other service tasks on a component of a wind turbine.

[0048] The service instruction generator may be implemented on an internet connected server. In this manner the service technician can via a client application, e.g. implemented on a mobile phone or a laptop computer, enter the query. The client application then forwards the query to the service instruction generator, which sends 2 the query to the trained natural language model that returns 3 the service instruction to the service instruction generator, which sends it to the client application for the service technician to receive either as a text message or as a voice message.

[0049] A query may be related to any service task at hand. As a simple example, the service technician may upon a service inspection find that a gasket is leaking hydraulic oil. A query may be formulated as: I need to change the gasket between component A and component B in the turbine Vxy-123, provide type number of relevant gasket as well as preferred steps to change the gasket.

[0050] This query is forwarded to the natural language model which in the manner as is known from e.g. ChatGPT may provide a service instruction of the form: The gasket serial number is GS- 456. To replace the gasket, perform the following steps A, B, C. With this service instruction, the service technician performs the service action, i.e. replacing the gasket.

[0051] The service task may be more complex, and in such a situation the service technician may query the service instruction generator a number of times before the actual service action is performed.

[0052] Figure 2 illustrates a flowchart representing general steps performed by the service instruction generator in accordance with embodiments of the present invention.

[0053] The service instruction generator is arranged with access 4 to a repository of service instructions RSI of specific wind turbine models, each service instructions in the repository of service instructions having a service instruction status tag ST.

[0054] The repository of service instructions may be a server comprising a hard disk on which service manuals are stored. Each service manual comprising service instructions in the form of information related to which specific wind turbine models it applies and a status tag, such as obsolete, or only valid for wind turbine model of a specific generation, or a specific variant.

[0055] The service instruction generator is also arranged with access 5 to a repository of fault entries RFE listing specific identified faults of specified wind turbines and associated repair actions done to mitigate the identified faults, each fault entry having a fault entry status tag. The repository of fault entries may be a server comprising a hard disk on which the fault entries of specific wind turbines are collected and stored. Each fault entry comprises associated repair actions done to mitigate the identified faults and a status tag, such as status standard repair or status custom repair.

[0056] At defined recurring time the service instruction generator is programmed to prepare a current text corpus. The current text corpus comprises a plurality of corpus elements used for training the natural language model. Rules are set up for the service instruction generator to at a recurring time build the current text corpus and train the natural language model with the current text corpus. The service instruction generator may comprise settings which define (e.g. user defined) selected service instruction status tags and selected fault entry status tags. With these selected status tags, the service instruction generator instructs 4, 5 the repositories to forward the repository entries and prepare 6 the current text corpus. The process may also be a dynamic process where the service instruction generator dynamically receives the repository entries and builds the current text corpus.

[0057] The current text corpus is at the defined recurring time forwarded 7 to the natural language model which is trained using the current text corpus. The trained natural language model is used to generate the service instruction until the next recurring time, after which the retrained on the then current text corpus natural language model is used to generate the service instruction, and so forth at each recurring time.

[0058] The service instruct! on / service action may in this manner be generated from a trained natural language model which is trained on a text corpus comprising only corpus elements with selected status tags.

[0059] The defined recurring time may in embodiments be a predefined recurring time interval. In an alternative embodiment, the defined recurring time is determined based on an amount of received corpus element(s) to be included in the current text corpus.

[0060] Training a natural language model on a new text corpus may be done in different ways. In embodiments the training may be based on a pre-trained natural language model, such as GPT or BERT model or another large-scale natural language model. This is known as finetuning or domain adaptation of the pre-trained language model.

[0061] In an embodiment, the current text corpus is created and maintained by the service instruction generator or a computer program instructed by the service instruction generator. The text corpus may be maintained in txt format or a pdf format. The current text corpus may then be tokenized by a tokenizer. A programmed training manager may based on the tokenized current text corpus re-train the natural language model in accordance with common practice within the field of natural language models.

[0062] In embodiments the current text corpus may comprise a number of further corpus elements originating from different sources. The different sources being sources of relevant information for performing the service action.

[0063] In an embodiment, the service instruction generator comprises a defined application programming interface (API) for receiving corpus elements to be included in the current text corpus. In this manner an appropriate source of corpus elements may be handled in an easy manner using the communication structure of the wind turbine under service. In general, an API may not be needed as other manners of receiving information to be included in the current text corpus is possible.

[0064] Figure 3 illustrates an example system for obtaining different types of corpus elements together with elements of a general wind turbine system.

[0065] Figure 3 illustrates elements of a wind turbine 30 together with functional elements of the service instruction generator SIG comprising an API and further elements that will be explained below. The wind turbine 30 includes a tower 32, a nacelle 33 disposed at the apex of the tower, and a rotor 34 operatively coupled to a generator (not shown) housed inside the nacelle. In addition to the generator, the nacelle houses miscellaneous components required for converting wind energy into electrical energy and various components needed to operate the wind turbine. The rotor 34 of the wind turbine includes a central hub 35 and a plurality of blades 36 that project outwardly from the central hub. Moreover, the wind turbine comprises a control system. The control system may be placed inside the nacelle or distributed at a number of locations inside (or externally to) the turbine and communicatively connected. The illustrated control system comprises a main controller 31 connected via an internal communication network 37 to a number of distributed controller nodes 38. The communication network further comprises a number of communication switches 39. Each distributed controller may be connected to a number of sensors 300.

[0066] The wind turbine 1 may be included among a collection of other wind turbines belonging to a wind power plant, also referred to as a wind farm or wind park, that serves as a power generating plant connected by transmission lines with a power grid.

[0067] The control system thus comprises a number of elements, including at least one controller with a processor and a memory, so that the processor is capable of executing computing tasks based on instructions stored in the memory. The control system may via a switch 301 and an external communication network be connected to an external computing system 302, such as a central controller, a supervising control and data acquisition system (SCAD A), or other type. In this manner may sensor readings, control data, etc., via the internal communication network and the external communication network be made available at the external computing system 302. A computing module in the external computing system may be programmed to make available data 305 to the service instruction generator SIG via the API.

[0068] Using the illustrated system of figure 3, the service instruction generator may be instructed to include sensor reading data from one or more sensors of the specific wind turbines as corpus elements of the current text corpus. As a very large number of sensor readings may be available in an operating wind turbine it may be beneficial to only include into the current text corpus the sensor readings that have been obtained in a specified period prior to preparing the current text corpus.

[0069] The API may provide a defined interface for providing corpus elements to be included in the current text corpus. Additionally, a number of further sources of information may via the API be included in the current text corpus and thereby made available to the natural language model. These sources of information are generally illustrated on figure 3 by elements 304. Advantageous information sources are mentioned in the following.

[0070] In an embodiment, a listing 304 of components of the specific wind turbines may be included into the current text corpus. The corpus elements listing components of the specific wind turbines may further comprises information related to at least the last service action performed on the components, and beneficially to all service actions that have been performed on the components of the wind turbines under service.

[0071] An automatic report setup may be available for a wind turbine, e.g. a report setup for providing operational data, status data, etc. to the wind turbine operator. Such report system may be based on a power BI system. Such report data may be included 304 via the API into the current text corpus.

[0072] Alarm logs, warning logs and other logs may be maintained in various databases, e.g. provided to a surveillance centre that monitors the operation of wind farms. Such log data may be included 304 via the API into the current text corpus.

[0073] A condition monitoring system may collect data and determine a state of the wind turbine. Such condition monitoring data may be included 304 via the API into the current text corpus.

[0074] Configuration data, e.g. for various controller components, sensors, etc. may be maintained in a configuration database. Such configuration data may be included 304 via the API into the current text corpus.

[0075] In an embodiment, the current text corpus may further be provided with corpus elements in the form of ad hoc selectable text of the specific wind turbines. For example, an email component 304 may be setup which provides an interface for forwarding specific emails, both ad hoc or according to given rules, via the API to the service instruction generator. The email component may e.g. be a mail server which is programmed to, if receiving an email on a specific email address to include such email into the current text corpus. Emails, or other text messages may be important sources of information for service repair. Email sent between two service technicians regarding a specific repair task may comprise relevant information.

[0076] The repository of service instructions of specific wind turbine models and the repository of fault entries may also be provided as a source of information 304 that can be accessed via the API. Figure 4 illustrates a flow diagram of an example text corpus manager TCM implemented as a computing module in the service instruction generator to prepare the current text corpus. The text corpus manager receives a corpus element CE, e.g. via the API as disclosed in figure 3.

[0077] In an embodiment the TCM compares 40 the received corpus element CE with the corpus elements of the existing text corpus and include 41 the received corpus element into the current text corpus if the received corpus element is not already included in the existing corpus material. If the corpus element is already included in the existing corpus material the corpus element is disregarded 42

[0078] In an embodiment, the corpus element comprises an assigned status tag ST. In such embodiment, the text corpus manager TCM further compares the status tag with a selected status tag SST and only includes 41 the corpus element into the current text corpus if the status tag of the received corpus element matches the selected status tag. The selected status tag may be a tag comprises in a list of status tags which are to be included in the current text corpus. If the status tag of the received corpus element does not match the selected status tag, the corpus element is disregarded 42.

[0079] Figure 4 illustrates a further embodiment that may be implemented as an addition or an alternative, and advantageously, both embodiments may be operated to prepare the current text corpus.

[0080] In the further embodiment the text corpus manager is implemented to remove 43 corpus elements from the current text corpus if the status tag of a given corpus element is different from a selected status tag SST. In an example implementation the text corpus manager receives 40 corpus elements of the existing text corpus, i.e. corpus elements already included in the current text corpus, and compares the corpus elements to a list 44 of corpus elements with changed status tags, if the changed status tag is no longer one of the selected status tags SST to be included in the current text corpus, the corpus element is removed 43 from the current text corpus, e.g. by deletion. In an embodiment a status tag of a service instructions of specific wind turbine model may have changed from valid to obsolete, and therefore removed from the current text corpus. The status tags may be provided manually or automatically using a tagging tool. The status tag may be provided in a manner so that it is embedded in the instruction or entry, or it may be provided as a specific meta data entry. For example, the instructions or entries may be stored in a structured format like XML or JSON where a data structure can be associated with a status tag. The status tags may also be stored in a tag database which based on an ID of the instruction or can be associated with a tag.

[0081] Example embodiments of the invention have been described for the purposes of illustration only, and not to limit the scope of the invention as defined in the accompanying claims.

Claims

CLAIMS1. A method of servicing a wind turbine using a computer implemented service instruction generator applying a trained natural language model; the method comprises: query the service instruction generator; wherein service instruction generator performs: accessing a repository of service instructions of specific wind turbine models, each service instructions in the repository of service instructions having a service instruction status tag; accessing a repository of fault entries listing specific identified faults of specified wind turbines and associated repair actions done to mitigate the identified faults, each fault entry having a fault entry status tag; at a defined recurring time prepare a current text corpus, wherein the current text corpus comprises a plurality of corpus elements, and wherein the plurality of corpus elements comprises the service instructions in the repository of service instructions having a selected service instruction status tag and the fault entries in the repository of fault entries having a selected fault entry status tag; train the natural language model on the current text corpus; generate a service instruction using the natural language model based on the query; and perform a service action on the wind turbine based on the service instruction.

2. The method according to claim 1, wherein the current text corpus further comprises corpus elements being sensor reading data from one or more sensors of the specific wind turbines.

3. The method according to claim 2, wherein the sensor readings have been obtained in a specified period prior to preparing the current text corpus.

4. The method according to any preceding claims, wherein the current text corpus further comprises corpus elements listing components of the specific wind turbines.

5. The method according to claim 4, wherein the corpus elements listing components of the specific wind turbines further comprises information related to at least the last service action performed on the components.

6. The method according to any preceding claims, wherein the current text corpus further comprises corpus elements of ad hoc selectable text of the specific wind turbines.

7. The method according to any of the preceding claims, wherein the service instruction generator comprises a defined application programming interface (API) for receiving corpus elements to be included in the current text corpus.

8. The method according to any of the preceding claims, wherein the service instruction generator prepares the current text corpus by comparing a received corpus element with the corpus elements of the existing text corpus and include the received corpus element into the current text corpus if the received corpus element is not included in the existing corpus material.

9. The method according to any preceding claims, wherein a corpus element comprises an assigned status tag, and wherein the service instruction generator prepares the current text corpus by identifying the status tag of the corpus elements of the existing text corpus and remove corpus elements from the current text corpus if the status tag of a given corpus element is different from a selected status tag.

10. The method according to any preceding claims, wherein the defined recurring time is a predefined recurring time interval between an hour and a week.

11. The method according to any of the claims 1-9, wherein the defined recurring time is determined based on an amount of received corpus element(s) to be included in the current text corpus.

12. The method according to any preceding claims, wherein the service instruction further comprises a listing of components to be used for performing the service action.

13. A computer program product comprising instructions which, when the program is executed by a computer, cause the computer to carry out the steps of the method of claims 1- 12.

14. A system for service of a wind turbine using a computer implemented service instruction generator applying a trained natural language model; the system comprises: a client device arranged for generating a query to be provided to the service instruction generator, and for receiving a generated service instruction from the service instruction generator; a processor; and a memory storing thereon executable instructions that, when executed, cause the processor to implement the service instruction generator to perform: accessing a repository of service instructions of specific wind turbine models, each service instructions in the repository of service instructions having a service instruction status tag; accessing a repository of fault entries listing specific identified faults of specified wind turbines and associated repair actions done to mitigate the identified faults, each fault entry having a fault entry status tag; at a defined recurring time prepare a current text corpus, wherein the current text corpus comprises a plurality of corpus elements, and wherein the plurality of corpus elements comprises the service instructions in the repository of service instructions having a selected service instruction status tag and the fault entries in the repository of fault entries having a selected fault entry status tag; train the natural language model on the current text corpus; generate the service instruction using the natural language model based on the query.

Images