A method and system for automatically extracting and structuring a wind turbine fault case into a database

Abstract

The application provides a method and system for automatically extracting and structuring a wind turbine fault case, pre-processes technical exchange data related to a wind turbine fault and inputs the data into a case fragment identification model to obtain a candidate case fragment; inputs the case identification fragment into a structured extraction model to identify a four-tuple entity, optimizes a relationship chain of the four-tuple entity based on a post-processing rule to obtain a plurality of four-tuple chain structures; evaluates the quality of the four-tuple chain structures based on field integrity, average entity confidence and negative keywords, converts a case related to the four-tuple chain structure into a standardized format, stores the case into a knowledge database or sends the case to an expert terminal for correction and verification, and uses the result as a new training set; and based on an incremental online learning strategy, regularly fine-tunes and iteratively optimizes the structured extraction model using the new training set. The application realizes automatic capture and reusable management of expert tacit knowledge, and solves the problem of difficult knowledge sedimentation in the prior art.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and intelligent operation and maintenance technology of wind turbine equipment, specifically to a method and system for automatically extracting and structurally storing wind turbine failure cases. Background Technology

[0002] In current wind turbine equipment operation and maintenance practices, valuable troubleshooting experience mostly exists in unstructured form in technical exchange data such as remote consultation records, expert WeChat group chats, and transcripts of voice conferences. Existing technologies mainly rely on manual reading, filtering, and input to transform the effective information in technical exchange data into structured knowledge base entries. Some systems use keyword matching for preliminary filtering. However, existing technologies have the following significant drawbacks: First, existing technologies rely on manual processing of technical data, which is not only inefficient and unable to cope with massive amounts of fragmented communication data, but also leads to a significant loss of tacit knowledge. Secondly, existing keyword matching methods cannot understand contextual semantics and are prone to missing complete case logic chains across sentences and speakers; Furthermore, existing technologies lack the ability to automatically identify the four-tuple relationship of "fault phenomenon-cause-measure-tool", and cannot generate structured knowledge that can be directly used for intelligent diagnosis. Moreover, the knowledge entry process lacks a quality assessment mechanism, making it easy for noisy data to be mixed into the knowledge base, affecting the reliability of subsequent AI systems.

[0003] Therefore, in view of the above-mentioned defects in the prior art, this application proposes a method and system for automatic extraction and structured storage of wind turbine failure cases. Summary of the Invention

[0004] This application addresses one or more technical deficiencies in the prior art by proposing the following technical solution.

[0005] Based on the first aspect of this application, a method for automatically extracting and structurally storing wind turbine failure cases is proposed, including: S1: Preprocess the technical communication data related to wind turbine failures, and input the preprocessed technical communication data into the case fragment recognition model to obtain candidate case fragments; S2: Input the case identification fragment into the structured extraction model to identify four-tuple entities. Based on post-processing rules, optimize the relational chain of the four-tuple entities to obtain multiple four-tuple chain structures. The four-tuple entities include the fault phenomenon, the fault cause, the handling measure, and the tool used. S3: The quality of the quadruple chain structure is evaluated based on field completeness, average entity confidence, and negative keywords. When the quality evaluation score is greater than or equal to a preset threshold, the cases related to the quadruple chain structure are converted into a standardized format and stored in the knowledge database. The formula for calculating the quality assessment score is as follows: Score = W1×A + W2×B + W3×C; Wherein, Score represents the quality assessment score, A represents the field completeness score, B represents the average score of entity confidence, C represents the negative keyword score, and W1, W2, and W3 represent the weights of the field completeness score, the average score of entity confidence, and the negative keyword score, respectively. S4: When the quality assessment score is less than the preset threshold, the case is sent to the expert for correction and verification, and the result is used as a new training set. Based on the incremental online learning strategy, the structured extraction model is fine-tuned periodically using the new training set, and the fine-tuned structured extraction model is iteratively optimized.

[0006] This application effectively reduces the loss of tacit knowledge, ensures the accuracy and reliability of standard cases, avoids the pollution of noisy data, and achieves continuous optimization of the model and self-improvement of the knowledge base by automating the process from raw technical exchange data to the storage of structured standard cases.

[0007] Furthermore, the preprocessing includes uniformly encoding the technical exchange data, removing webpage tags and special symbols, separating the dialogue stream according to the speaker's identity, and marking each speech with a timestamp and source identifier; The technical exchange data includes chat logs, transcripts of remote video conferences, expert audio recordings, and email content.

[0008] This step preserves the context and timing information of the technical exchange data, which is beneficial for subsequent fragment recognition and relationship extraction.

[0009] Furthermore, the case fragment recognition model is a BERT-based binary classification model deployed on the edge. A context window consisting of a single statement or multiple consecutive statements is input into the case fragment recognition model to obtain candidate case fragments containing complete fault handling logic. By back-matching technical exchange data with standard cases already in the database, the matched technical exchange data is marked as positive samples, and the remaining unmatched technical exchange data is marked as negative samples. The case fragment recognition model is then trained using the positive and negative samples.

[0010] Deploying the case fragment recognition model at the edge enables real-time processing of technical exchange data and achieves data localization, reducing network transmission pressure and data leakage risks, and improving system security and response speed.

[0011] Furthermore, the structured extraction model is a BERT-CRF sequence labeling model deployed in a central cloud. The BERT-CRF sequence labeling model is pre-trained using domain-adaptive methods. Based on the general Chinese BERT model, a large amount of professional text is used to perform secondary pre-training on the BERT-CRF model.

[0012] Domain-adaptive pre-training enables structured extraction models to better understand the terminology and context of the wind turbine industry, improving the accuracy of quadruple entity recognition. Deploying structured extraction models in a central cloud allows for more powerful computing to handle complex sequence labeling tasks.

[0013] Furthermore, the relationship chain of the quadruple entities is optimized based on post-processing rules. Specifically, if a causal correlation word appears in the technical exchange data, the entity following the word is forcibly classified as the fault cause (Reason). If the verb "tool use" appears in the technical exchange data, the object entity of the verb will be forcibly classified as "Tool". If the technical exchange data contains descriptions of abnormal equipment status, specific abnormal equipment, and abnormal parameters, then the description shall be classified as a fault phenomenon (Phenomenon). If a verb-object phrase containing a specific part and the object of operation appears in the technical exchange data, then the verb-object phrase is classified as a processing measure (Action).

[0014] This step ensures that key information is not overlooked, maintains high entity extraction stability in conversational and fragmented text, and improves the logical consistency and interpretability of structured extraction.

[0015] Furthermore, the probability value of each word in the quadruple entity output by the structured extraction model CRF layer is predicted as the correct label is calculated, and the average probability value of all words in the same quadruple entity is calculated to obtain the average confidence score of the entity.

[0016] Furthermore, the field completeness is the completeness of the four-tuple entity; the more complete the entity tokens, the higher the field completeness score. The negative keywords are words related to unresolved or recurring problems. The more negative keywords there are, the lower the score for each negative keyword.

[0017] Furthermore, the step of converting the cases related to the quadruple chain structure into a standardized format and storing them in the knowledge database includes converting the relevant cases into standardized JSON objects and writing them into the Neo4j knowledge graph and relational database. The standardized JSON objects contain entity content, device model, first appearance time, and source ID information. Before storing standard cases into the knowledge database, the four-element chain structure is standardized to generate a unique content fingerprint. If the fingerprint is completely consistent with the fingerprint of an existing case in the knowledge database, the standard case is a duplicate case. The system automatically discards the standard case and appends the source ID information to the source list of existing cases. If the fingerprint is inconsistent with the fingerprint of an existing case, the standard case is stored as a new case and waits for experts to merge or remove duplicates.

[0018] Based on the second aspect of this application, a system for automatically extracting and structurally storing wind turbine failure cases is also proposed, including: Case Fragment Recognition Module: Preprocesses technical communication data related to wind turbine failures and inputs the preprocessed technical communication data into the case fragment recognition model to obtain candidate case fragments; The structured extraction module inputs the case recognition fragments into the structured extraction model to identify four-tuple entities. Based on post-processing rules, it optimizes the relational chains of the four-tuple entities to obtain multiple four-tuple chain structures. The four-tuple entities include the fault phenomenon, the fault cause, the handling measure, and the tool used. The database entry module performs a quality assessment on the quadruple chain structure based on field completeness, average entity confidence, and negative keywords. When the quality assessment score is greater than or equal to a preset threshold, the cases related to the quadruple chain structure are converted into a standardized format and stored in the knowledge database. The formula for calculating the quality assessment score is as follows: Score = W1×A + W2×B + W3×C; Wherein, Score represents the quality assessment score, A represents the field completeness score, B represents the average score of entity confidence, C represents the negative keyword score, and W1, W2, and W3 represent the weights of the field completeness score, the average score of entity confidence, and the negative keyword score, respectively. Optimize the feedback module: When the quality assessment score is less than the preset threshold, the case is sent to the expert for correction and verification, and the result is used as a new training set. Based on the incremental online learning strategy, the structured extraction model is fine-tuned periodically using the new training set, and the fine-tuned structured extraction model is iteratively optimized.

[0019] Based on a third aspect of this application, a computer program product is also proposed, having one or more computer programs thereon that, when executed by a computer processor, implement the method described above.

[0020] The technical advantages of this application are as follows: This application replaces manual operation with an end-to-end NLP pipeline, performs domain-adaptive pre-training on the structured extraction model and incorporates a built-in quality scoring mechanism, and provides a method and system for automatic extraction and structured storage of wind turbine failure cases. It can automatically, accurately and efficiently extract structured failure cases from wind turbine technical exchange data, effectively filter out low-confidence case fragments, realize the automated capture and reusable management of expert tacit knowledge, and solve the problems of manual extraction, high omission rate, low degree of structure and difficulty in knowledge accumulation in the prior art. Attached Figure Description

[0021] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings.

[0022] Figure 1 This is a flowchart of a method for automatically extracting and structurally storing wind turbine failure cases according to an embodiment of this application.

[0023] Figure 2 This is a system module diagram of automatic extraction and structured storage of wind turbine failure cases according to an embodiment of this application.

[0024] Figure 3 This is a schematic diagram of the structure of a computer system suitable for implementing the electronic devices of the present application embodiments. Detailed Implementation

[0025] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0026] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0027] Figure 1 This application illustrates a method for automatically extracting and structurally storing wind turbine failure cases, including: S1: Preprocess the technical communication data related to wind turbine failures, and input the preprocessed technical communication data into the case fragment recognition model to obtain candidate case fragments; S2: Input the case identification fragment into the structured extraction model to identify four-tuple entities. Based on post-processing rules, optimize the relational chain of the four-tuple entities to obtain multiple four-tuple chain structures. The four-tuple entities include the fault phenomenon, the fault cause, the handling measure, and the tool used. S3: The quality of the quadruple chain structure is evaluated based on field completeness, average entity confidence, and negative keywords. When the quality evaluation score is greater than or equal to a preset threshold, the cases related to the quadruple chain structure are converted into a standardized format and stored in the knowledge database. The formula for calculating the quality assessment score is as follows: Score = W1×A + W2×B + W3×C; Wherein, Score represents the quality assessment score, A represents the field completeness score, B represents the average score of entity confidence, C represents the negative keyword score, and W1, W2, and W3 represent the weights of the field completeness score, the average score of entity confidence, and the negative keyword score, respectively. S4: When the quality assessment score is less than the preset threshold, the case is sent to the expert for correction and verification, and the result is used as a new training set. Based on the incremental online learning strategy, the structured extraction model is fine-tuned periodically using the new training set, and the fine-tuned structured extraction model is iteratively optimized.

[0028] It should be noted that this application effectively reduces the loss of tacit knowledge, ensures the accuracy and reliability of standard cases, avoids the pollution of noisy data, and achieves continuous optimization of the model and self-improvement of the knowledge base by automating the process from raw technical exchange data to the storage of structured standard cases.

[0029] It should be noted that the preprocessing includes uniformly encoding the technical exchange data, removing web page tags and special symbols, separating the dialogue stream according to the speaker's identity, and marking each speech with a timestamp and source identifier; The technical exchange data includes chat logs, transcripts of remote video conferences, expert audio recordings, and email content.

[0030] It should be noted that this step preserves the context and timing information of the technical exchange data, which is beneficial for subsequent fragment recognition and relationship extraction.

[0031] It should be noted that the case fragment recognition model is a BERT-based binary classification model deployed on the edge. The case fragment recognition model is input into a context window consisting of a single statement or multiple consecutive statements to obtain candidate case fragments containing complete fault handling logic. By back-matching technical exchange data with standard cases already in the database, the matched technical exchange data is marked as positive samples, and the remaining unmatched technical exchange data is marked as negative samples. The case fragment recognition model is then trained using the positive and negative samples.

[0032] It should be noted that deploying the case fragment recognition model on the edge allows for real-time processing of technical exchange data and enables data localization, reducing network transmission pressure and the risk of data leakage, and improving system security and response speed.

[0033] It should be noted that the structured extraction model is a BERT-CRF sequence labeling model deployed on a central cloud. The BERT-CRF sequence labeling model is pre-trained using domain-adaptive methods. Based on the general Chinese BERT model, a large amount of professional text is used to perform secondary pre-training on the BERT-CRF model.

[0034] It should be noted that domain-adaptive pre-training enables the structured extraction model to better understand the terminology and context of the wind turbine industry, improve the accuracy of quadruple entity recognition, and deploy the structured extraction model on the central cloud to handle complex sequence labeling tasks with stronger computing power.

[0035] It should be noted that the relationship chain optimization of the quadruple entities is based on the post-processing rules. Specifically, if a causal correlation word appears in the technical exchange data, the entity following that word will be forcibly classified as the fault cause (Reason). If the verb "tool use" appears in the technical exchange data, the object entity of the verb will be forcibly classified as "Tool". If the technical exchange data contains descriptions of abnormal equipment status, specific abnormal equipment, and abnormal parameters, then the description shall be classified as a fault phenomenon (Phenomenon). If a verb-object phrase containing a specific part and the object of operation appears in the technical exchange data, then the verb-object phrase is classified as a processing measure (Action).

[0036] It should be noted that this step ensures that key information is not overlooked, maintains high entity extraction stability in colloquial and fragmented text, and improves the logical consistency and interpretability of structured extraction.

[0037] It should be noted that the probability value of each word in the quadruple entity output by the CRF layer of the structured extraction model is predicted as the correct label is calculated, and the average of the probability values ​​of all words in the same quadruple entity is calculated to obtain the average confidence score of the entity.

[0038] It should be noted that the field completeness refers to the completeness of the four-tuple entity; the more complete the entity tokens, the higher the field completeness score. The negative keywords are words related to unresolved or recurring problems. The more negative keywords there are, the lower the score for each negative keyword.

[0039] It should be noted that converting the cases related to the quadruple chain structure into a standardized format and storing them in the knowledge database includes converting the relevant cases into standardized JSON objects and writing them into the Neo4j knowledge graph and relational database. The standardized JSON objects contain entity content, device model, first appearance time, and source ID information. Before storing standard cases into the knowledge database, the four-element chain structure is standardized to generate a unique content fingerprint. If the fingerprint is completely consistent with the fingerprint of an existing case in the knowledge database, the standard case is a duplicate case. The system automatically discards the standard case and appends the source ID information to the source list of existing cases. If the fingerprint is inconsistent with the fingerprint of an existing case, the standard case is stored as a new case and waits for experts to merge or remove duplicates.

[0040] It should be noted that this application replaces manual operation with an end-to-end NLP pipeline, performs domain-adaptive pre-training on the structured extraction model, and incorporates a built-in quality scoring mechanism. It provides a method and system for automatically extracting and structured storage of wind turbine failure cases. It can automatically, accurately, and efficiently extract structured failure cases from wind turbine technical exchange data, effectively filter out low-confidence case fragments, realize the automated capture and reusable management of expert tacit knowledge, and solve the problems of manual extraction, high omission rate, low degree of structure, and difficulty in knowledge accumulation in the prior art.

[0041] The following is for reference. Figure 2 The system demonstrates an automatic extraction and structured storage system for wind turbine failure cases, including a case fragment recognition module 201, a structure extraction module 202, a storage module 203, and an optimization feedback module 204.

[0042] In a specific embodiment, the case fragment recognition module 201 is configured to: preprocess the technical communication data related to wind turbine failure, and input the preprocessed technical communication data into the case fragment recognition model to obtain candidate case fragments.

[0043] In a specific embodiment, the structured extraction module 202 is configured to: input the case recognition fragment into the structured extraction model to identify four-tuple entities, optimize the relationship chain of the four-tuple entities based on post-processing rules, and obtain multiple four-tuple chain structures. The four-tuple entities include the fault phenomenon, the fault cause, the handling measure, and the tool used.

[0044] In a specific embodiment, the database entry module 203 is configured to: perform a quality assessment on the quadruple chain structure based on field completeness, average entity confidence, and negative keywords; when the quality assessment score is greater than or equal to a preset threshold, convert the cases related to the quadruple chain structure into a standardized format and store them in the knowledge database. The formula for calculating the quality assessment score is as follows: Score = W1×A + W2×B + W3×C; Wherein, Score represents the quality assessment score, A represents the field completeness score, B represents the average score of entity confidence, C represents the negative keyword score, and W1, W2, and W3 represent the weights of the field completeness score, the average score of entity confidence, and the negative keyword score, respectively.

[0045] In a specific embodiment, the optimization feedback module 204 is configured to: when the quality assessment score is less than a preset threshold, send the case to the expert terminal for correction and verification and use the result as a new training set; based on an incremental online learning strategy, periodically use the new training set to fine-tune the structured extraction model and iteratively optimize the fine-tuned structured extraction model.

[0046] It should be noted that the preprocessing includes uniformly encoding the technical exchange data, removing web page tags and special symbols, separating the dialogue stream according to the speaker's identity, and marking each speech with a timestamp and source identifier; The technical exchange data includes chat logs, transcripts of remote video conferences, expert audio recordings, and email content.

[0047] It should be noted that the case fragment recognition model is a BERT-based binary classification model deployed on the edge. The case fragment recognition model is input into a context window consisting of a single statement or multiple consecutive statements to obtain candidate case fragments containing complete fault handling logic. By back-matching technical exchange data with standard cases already in the database, the matched technical exchange data is marked as positive samples, and the remaining unmatched technical exchange data is marked as negative samples. The case fragment recognition model is then trained using the positive and negative samples.

[0048] In a specific embodiment, the existing standard case library within the company, which is manually sorted and stored in the database by experts, is used as the "gold standard". The key information in each standard case (such as specific fault phenomena, equipment models, and handling measures) is used as keywords for full-text retrieval in a large amount of historical technical communication data. Whenever multiple key elements of a certain standard case are simultaneously included in a continuous speech segment, for example, both "abnormal noise in the pitch motor" and "replace the encoder" are mentioned, then this speech segment is automatically marked as a positive sample, and the remaining speech segments that are not matched or only contain sporadic irrelevant information are marked as negative samples. In this way, a large-scale training dataset can be quickly constructed without manual annotation item by item.

[0049] In a specific embodiment, the BERT binary classification model introduces a context window sliding verification mechanism. For a candidate case segment determined to contain a complete fault handling, the system automatically expands several adjacent conversation speeches before and after it to form a larger context window and inputs it into the model again for review. When this candidate case segment is determined to be a positive sample under both the original window and the expanded window, it is the finally confirmed valid case candidate segment.

[0050] It should be noted that deploying the case segment recognition model on the edge side (such as the local server of a wind farm) can process a large amount of on-site generated technical communication data in real time without uploading sensitive data to the cloud, which not only reduces the network bandwidth pressure but also ensures the security and privacy of the enterprise's technical data.

[0051] It should be noted that the structured extraction model is a BERT-CRF sequence annotation model deployed on the central cloud. Domain adaptation pre-training is performed on the BERT-CRF sequence annotation model, and a large amount of professional text is used to perform secondary pre-training on the BERT-CRF model based on the general Chinese BERT model.

[0052] In a specific embodiment, based on the general Chinese BERT model, a large amount of professional text in the wind power field is used to perform secondary pre-training on BERT-CRF, including fan technical manuals, maintenance work orders, industry standards, and historical communication records. In this process, the BERT-CRF model can deeply learn the professional terms, expression habits, and context features in the wind power field. For example, "overspeed" specifically refers to "overspeed" in the wind power field rather than simply a fast speed in general. The trained BERT-CRF model can significantly improve the entity recognition accuracy in the subsequent recognition of quadruple entities. Deploying it on the central cloud can utilize stronger computing power to handle complex sequence annotation tasks.

[0053] It should be noted that the relationship chain optimization of the quadruple entities is based on the post-processing rules. Specifically, if a causal correlation word appears in the technical exchange data, the entity following that word will be forcibly classified as the fault cause (Reason). If the verb "tool use" appears in the technical exchange data, the object entity of the verb will be forcibly classified as "Tool". If the technical exchange data contains descriptions of abnormal equipment status, specific abnormal equipment, and abnormal parameters, then the description shall be classified as a fault phenomenon (Phenomenon). If a verb-object phrase containing a specific part and the object of operation appears in the technical exchange data, then the verb-object phrase is classified as a processing measure (Action).

[0054] In a specific embodiment, if the word "due to" or "cause" appears in the technical communication data, the entity following the word is classified as the cause of the failure. If the technical exchange data contains the words "suggest", "need", or "use", then the object of the verb will be classified as "Tool". If the technical communication data contains words such as "error", "alarm", "abnormal", "unable" or "failed" that describe abnormal equipment status, and includes specific equipment or parameters, such as "pitch communication interrupted" or "gearbox oil temperature too high", then the description is classified as a fault phenomenon (Phenomenon). If the technical exchange data contains words such as "inspect", "replace", "tighten", "restart", "calibrate", or "clean", and the object of these verbs is a specific part or operation object, such as "replace the main control module" or "tighten the tower bolts", then it is treated as a processing action.

[0055] It should be noted that the probability value of each word in the quadruple entity output by the CRF layer of the structured extraction model is predicted as the correct label is calculated, and the average of the probability values ​​of all words in the same quadruple entity is calculated to obtain the average confidence score of the entity.

[0056] It should be noted that the field completeness refers to the completeness of the four-tuple entity; the more complete the entity tokens, the higher the field completeness score. The negative keywords are words related to unresolved or recurring problems. The more negative keywords there are, the lower the score for each negative keyword.

[0057] In a specific embodiment, the quadruple entity is {[Phenomenon]: "Generator bearing temperature continuously exceeds 90℃", [Reason]: "Grease aging and failure", [Action]: "Stop the machine and replace the grease"}. A tool is used to assess its quality. Assuming the maximum field completeness score is 1, deducting 0.25 points for each missing item, the field completeness score is 0.75. The maximum negative keyword score is 1, deducting 0.3 points for each negative keyword. The average entity confidence score output by the BERT-CRF model is 0.88, with a preset threshold of 0.7. No points are deducted for not containing negative keywords, and the negative keyword score is 1. Assuming equal weights for all three, the maximum total score is 3. Converted to a percentage, the quality assessment score is 0.88, which is greater than the preset threshold of 0.7. Therefore, the case meets the standard and can be included in the database. The system periodically pushes cases with entity confidence scores between 0.5 and 0.7 to the expert end in the form of a structured form. The expert end corrects the content of the four-tuple and directly feeds it back into the training set as a new training sample to fine-tune the BERT-CRF model. After the fine-tuned model passes internal verification and meets the performance standards, it replaces the old model online and is used to process subsequent new data, ensuring that it can continuously learn expert knowledge and on-site expression habits and continuously improve performance.

[0058] It should be noted that converting the cases related to the quadruple chain structure into a standardized format and storing them in the knowledge database includes converting the relevant cases into standardized JSON objects and writing them into the Neo4j knowledge graph and relational database. The standardized JSON objects contain entity content, device model, first appearance time, and source ID information. Before storing standard cases into the knowledge database, the four-element chain structure is standardized to generate a unique content fingerprint. If the fingerprint is completely consistent with the fingerprint of an existing case in the knowledge database, the standard case is a duplicate case. The system automatically discards the standard case and appends the source ID information to the source list of existing cases. If the fingerprint is inconsistent with the fingerprint of an existing case, the standard case is stored as a new case and waits for experts to merge or remove duplicates.

[0059] In a specific implementation, the quadruple chain structure (P,R,A,T) is standardized (e.g., synonym unification or unit unification) to generate a unique content fingerprint. If the content fingerprint is completely consistent with the fingerprint of an existing case in the knowledge base, it is considered an important case. The system discards the new case and appends the source ID information to the source list of existing cases to track the source of information. If the content fingerprint is inconsistent with the fingerprint of an existing case but highly similar (e.g., only the wording of the measures is different), the standard case is added to the database as a new case and awaits merging or deduplication by experts.

[0060] It should be noted that this application supports contextual understanding across statements and sentences, can completely reconstruct the fault handling logic chain, output a standard four-tuple structure, which can be directly applied to the construction of knowledge graphs and intelligent question answering systems. It has a built-in quality scoring mechanism that can effectively filter credibility fragments, ensure the purity of the knowledge base, automatically capture and structure the tacit knowledge of experts, and significantly improve the efficiency of knowledge asset accumulation.

[0061] The following is for reference. Figure 3 It shows a schematic diagram of the structure of a computer system suitable for implementing the electronic device of the present application. Figure 3 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0062] like Figure 3 As shown, the computer system includes a central processing unit (CPU) 301, which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) 302 or programs loaded from storage section 308 into random access memory (RAM) 303. The RAM 303 also stores various programs and data required for system operation. The CPU 301, ROM 302, and RAM 303 are interconnected via bus 304. An input / output (I / O) interface 305 is also connected to bus 304.

[0063] The following components are connected to I / O interface 305: an input section 306 including a keyboard, mouse, etc.; an output section 307 including a liquid crystal display (LCD) and speakers, etc.; a storage section 308 including a hard disk, etc.; and a communication section 309 including a network interface card such as a LAN card and a modem, etc. The communication section 309 performs communication processing via a network such as the Internet. A drive 310 is also connected to I / O interface 305 as needed. A removable medium 311, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 310 as needed so that computer programs read from it can be installed into storage section 308 as needed.

[0064] Specifically, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable storage medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 309, and / or installed from removable medium 311. When the computer program is executed by central processing unit (CPU) 301, it performs the functions defined in the methods of this application. It should be noted that the computer-readable storage medium of this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. The computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium may also be any computer-readable storage medium other than a computer-readable storage medium that can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. Program code contained on a computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0065] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof. Programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, and C++—as well as conventional procedural programming languages—such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0066] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0067] The modules described in the embodiments of this application can be implemented in software or in hardware.

[0068] In another aspect, this application also provides a computer-readable storage medium, which may be included in the electronic device described in the above embodiments; or it may exist independently and not assembled into the electronic device. The aforementioned computer-readable storage medium carries one or more programs. When the electronic device executes one or more of these programs, the electronic device performs the following actions: preprocesses technical communication data related to wind turbine faults and inputs the preprocessed technical communication data into a case fragment recognition model to obtain candidate case fragments; inputs the case recognition fragments into a structured extraction model to identify four-tuple entities, optimizes the relationship chain of the four-tuple entities based on post-processing rules, and obtains multiple four-tuple chain structures, wherein the four-tuple entities include a fault phenomenon (Phenomenon), a fault cause (Reason), a handling measure (Action), and a tool (Tool); performs a quality assessment of the four-tuple chain structures based on field completeness, average entity confidence, and negative keywords; when the quality assessment score is greater than or equal to a preset threshold, converts the cases related to the four-tuple chain structure into a standardized format and stores them in a knowledge database; when the quality assessment score is less than the preset threshold, sends the cases to an expert for correction and verification, uses the results as a new training set, and periodically fine-tunes the structured extraction model using the new training set based on an incremental online learning strategy, and iteratively optimizes the fine-tuned structured extraction model.

[0069] Finally, it should be noted that the above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A method for automatically extracting and structurally storing wind turbine failure cases, characterized in that, include: S1: Preprocess the technical communication data related to wind turbine failures, and input the preprocessed technical communication data into the case fragment recognition model to obtain candidate case fragments; S2: Input the case identification fragment into the structured extraction model to identify four-tuple entities. Based on post-processing rules, optimize the relational chain of the four-tuple entities to obtain multiple four-tuple chain structures. The four-tuple entities include the fault phenomenon, the fault cause, the handling measure, and the tool used. S3: The quality of the quadruple chain structure is evaluated based on field completeness, average entity confidence, and negative keywords. When the quality evaluation score is greater than or equal to a preset threshold, the cases related to the quadruple chain structure are converted into a standardized format and stored in the knowledge database. The formula for calculating the quality assessment score is: Score = W1 × A + W2 × B + W3 × C; Wherein, Score represents the quality assessment score, A represents the field completeness score, B represents the average score of entity confidence, C represents the negative keyword score, and W1, W2, and W3 represent the weights of the field completeness score, the average score of entity confidence, and the negative keyword score, respectively. S4: When the quality assessment score is less than the preset threshold, the case is sent to the expert for correction and verification, and the result is used as a new training set. Based on the incremental online learning strategy, the structured extraction model is fine-tuned periodically using the new training set, and the fine-tuned structured extraction model is iteratively optimized.

2. The method according to claim 1, characterized in that, The preprocessing includes uniformly encoding the technical exchange data, removing web page tags and special symbols, separating the dialogue stream according to the speaker's identity, and marking each speech with a timestamp and source identifier; The technical exchange data includes chat logs, transcripts of remote video conferences, expert audio recordings, and email content.

3. The method according to claim 1, characterized in that, The case fragment recognition model is a BERT-based binary classification model deployed on the edge. A context window consisting of a single statement or multiple consecutive statements is input into the case fragment recognition model to obtain candidate case fragments containing complete fault handling logic. By back-matching technical exchange data with standard cases already in the database, the matched technical exchange data is marked as positive samples, and the remaining unmatched technical exchange data is marked as negative samples. The case fragment recognition model is then trained using the positive and negative samples.

4. The method according to claim 1, characterized in that, The structured extraction model is a BERT-CRF sequence labeling model deployed in a central cloud. The BERT-CRF sequence labeling model is pre-trained using domain-adaptive methods. Based on the general Chinese BERT model, a large amount of professional text is used to perform secondary pre-training on the BERT-CRF model.

5. The method according to claim 1, characterized in that, The relationship chain of the quadruple entities is optimized based on post-processing rules. Specifically, if a causal correlation word appears in the technical exchange data, the entity following the word is forcibly classified as the fault cause (Reason). If the verb "tool use" appears in the technical exchange data, the object entity of the verb will be forcibly classified as "Tool". If the technical exchange data contains descriptions of abnormal equipment status, specific abnormal equipment, and abnormal parameters, then the description shall be classified as a fault phenomenon (Phenomenon). If a verb-object phrase containing a specific part and the object of operation appears in the technical exchange data, then the verb-object phrase is classified as a processing measure (Action).

6. The method according to claim 1, characterized in that, The probability value of each word in the quadruple entity output by the CRF layer of the structured extraction model is predicted as the correct label is calculated, and the average probability value of all words in the same quadruple entity is calculated to obtain the average confidence score of the entity.

7. The method according to claim 1, characterized in that, The field completeness score refers to the completeness of the four-tuple entity; the more complete the entity tokens, the higher the field completeness score. The negative keywords are words related to unresolved or recurring problems. The more negative keywords there are, the lower the score for each negative keyword.

8. The method according to claim 1, characterized in that, The step of converting the cases related to the quadruple chain structure into a standardized format and storing them in the knowledge database includes converting the relevant cases into standardized JSON objects and writing them into the Neo4j knowledge graph and relational database. The standardized JSON objects contain entity content, device model, first appearance time and source ID information. Before storing standard cases into the knowledge database, the four-element chain structure is standardized to generate a unique content fingerprint. If the fingerprint is completely consistent with the fingerprint of an existing case in the knowledge database, the standard case is a duplicate case. The system automatically discards the standard case and appends the source ID information to the source list of existing cases. If the fingerprint is inconsistent with the fingerprint of an existing case, the standard case is stored as a new case and waits for experts to merge or remove duplicates.

9. A system for automatically extracting and structurally storing wind turbine failure cases, characterized in that, include: Case Fragment Recognition Module: Preprocesses technical communication data related to wind turbine failures and inputs the preprocessed technical communication data into the case fragment recognition model to obtain candidate case fragments; The structured extraction module inputs the case recognition fragments into the structured extraction model to identify four-tuple entities. Based on post-processing rules, it optimizes the relational chains of the four-tuple entities to obtain multiple four-tuple chain structures. The four-tuple entities include the fault phenomenon, the fault cause, the handling measure, and the tool used. The database entry module performs a quality assessment on the quadruple chain structure based on field completeness, average entity confidence, and negative keywords. When the quality assessment score is greater than or equal to a preset threshold, the cases related to the quadruple chain structure are converted into a standardized format and stored in the knowledge database. The formula for calculating the quality assessment score is: Score = W1 × A + W2 × B + W3 × C; Wherein, Score represents the quality assessment score, A represents the field completeness score, B represents the average score of entity confidence, C represents the negative keyword score, and W1, W2, and W3 represent the weights of the field completeness score, the average score of entity confidence, and the negative keyword score, respectively. Optimize the feedback module: When the quality assessment score is less than the preset threshold, the case is sent to the expert for correction and verification, and the result is used as a new training set. Based on the incremental online learning strategy, the structured extraction model is fine-tuned periodically using the new training set, and the fine-tuned structured extraction model is iteratively optimized.

10. A computer program product having one or more computer programs thereon, characterized in that, When the computer program is executed by a computer processor, the method described in any one of claims 1-8 is performed.

Images