A zero-shot domain dictionary automatic construction method and system for wind power generation fault diagnosis

Abstract

This paper discloses a method and system for zero-shot automatic construction of a domain dictionary for wind power fault diagnosis. The method includes: first, receiving raw text from wind power operation and maintenance, extracting n-gram segments, and initially screening high-quality candidate words based on mutual information and left / right neighbor entropy; second, calculating and fusing the three-dimensional features of the candidate words—domain concentration, domain specificity, and life curve slope—to obtain a comprehensive domain score; third, automatically determining an adaptive threshold using the maximum inter-class variance method to classify candidate words into pseudo-positive and negative samples to construct a pseudo-label training set; and finally, using the pseudo-labels to perform supervised training and self-training iterations on a character-level convolutional neural network to generate a domain dictionary. This invention solves the problems of reliance on manual annotation and cold-start difficulties in traditional methods by introducing temporal evolution features and the Otsu automatic thresholding mechanism, achieving zero-shot fully automatic mining of wind power terminology.

Description

Technical Field

[0001] This invention relates to the technical fields of artificial intelligence and wind power operation and maintenance, and in particular to a method and system for automatic construction of a zero-sample domain dictionary for wind power fault diagnosis. Background Technology

[0002] With the rapid development of wind power technology, wind farms are expanding in scale. During the daily operation and maintenance of wind turbines, massive amounts of unstructured text data have accumulated, including maintenance work orders, technical bulletins, operation and maintenance logs, and fault descriptions. This text data contains rich information on equipment status, fault modes, and maintenance experience, forming a crucial foundation for intelligent fault diagnosis, knowledge graph construction, and automated operation and maintenance of wind turbines.

[0003] Current dictionary construction in the wind power field mainly relies on the following two methods, but both have significant drawbacks: The most common method currently is to rely on human experts to build the terminology. Engineers familiar with the wind turbine structure manually extract terminology from standard documents or maintenance records and compile them into a glossary. The disadvantages are: extremely high labor costs and low efficiency. Experts struggle to cover massive amounts of historical data and are prone to overlooking non-standard expressions. Furthermore, manual compilation is time-consuming and cannot meet the demands of rapid iteration.

[0004] Some existing technologies attempt to mine candidate words from corpora using statistical indicators such as TF-IDF (Term Frequency-Inverse Document Frequency), mutual information (MI), or left and right entropy. However, these methods suffer from several drawbacks: First, they lack domain-specific discrimination capabilities. Traditional statistical methods primarily measure the cohesion or boundary freedom within words, failing to effectively distinguish between "general high-frequency words" and truly valuable "domain-specific terms" for fault diagnosis. Second, they face difficulties in cold starts and rely heavily on manual intervention. These methods typically require pre-setting complex thresholds or providing an initial "seed dictionary" as a benchmark. In "zero-sample" scenarios where no dictionary has been accumulated at the project's inception, work often becomes impossible, creating a vicious cycle of "no dictionary leading to difficulty in processing, and difficulty in processing leading to a lack of knowledge." Third, they lack dynamic evolution awareness. Existing dictionary construction methods are mostly based on static corpus analysis; however, wind power technology updates rapidly, with new turbine models and fault types emerging suddenly over time. Existing technologies struggle to capture the evolutionary trends of vocabulary over time, resulting in lagging dictionary updates and an inability to adapt to the evolution of knowledge.

[0005] Furthermore, wind farms are typically located in remote areas with limited network bandwidth. Traditional centralized processing methods require uploading large amounts of raw text data to the cloud, which not only increases transmission overhead but may also raise data privacy concerns. Summary of the Invention

[0006] To address the aforementioned technical problems in the existing technology, this invention proposes a zero-sample domain dictionary automatic construction method and system for wind power generation fault diagnosis, thereby solving the above-mentioned technical problems.

[0007] According to a first aspect of the present invention, a method for automatic construction of a zero-sample domain dictionary for fault diagnosis in wind power generation is proposed, comprising: S1: Receive raw text data in the field of wind power operation and maintenance, preprocess the raw text data and extract n-gram segments as initial candidate words, calculate the mutual information and left and right neighboring word entropy of each initial candidate word, and retain high-quality candidate words based on preset statistical screening conditions. S2: For each high-quality candidate word obtained in step S1, calculate the three-dimensional feature index and perform weighted fusion to obtain the comprehensive domain score of the candidate word; S3: Based on the comprehensive domain score distribution of high-quality candidate words, the optimal adaptive threshold for distinguishing domain words from non-domain words is automatically calculated using the maximum inter-class variance method. Based on the adaptive threshold θ, high-quality candidate words are automatically classified to construct a pseudo-label training set. S4: Construct a character-level convolutional neural network model and supervise its training using a pseudo-label training set; use the trained character-level convolutional neural network model to predict all high-quality candidate words from step S1, select words with prediction confidence higher than the preset iteration threshold to expand the pseudo-label training set, and use the expanded training set to train the model for the next round of iterations until the preset stopping condition is reached. S5: Output the top-ranked words in the final prediction confidence score, and generate a dictionary for wind power fault diagnosis that includes word names, confidence scores, and part-of-speech tags.

[0008] In some specific embodiments, the preprocessing of the original text data in S1 specifically includes: receiving wind turbine maintenance work orders, technical notices and operation and maintenance logs, and unifying the text encoding to UTF-8 format; using the SimHash algorithm to deduplicate the text; filtering out text segments with character lengths less than a preset lower limit or greater than a preset upper limit; and segmenting the text into independent sentences according to punctuation marks.

[0009] In some specific embodiments, the three-dimensional feature indicators in S2 include the Domain Concentration (DR) value, the Inverse Document Frequency Difference (ΔIDF), and the slope of the life curve. The formula for calculating the Domain Concentration (DR) value is as follows: ,in, The frequency of candidate words in the corpus of wind power-related terms. The term frequency of the candidate word in the general corpus; the formula for calculating the inverse document frequency difference ΔIDF is: ,in, Inverse document frequencies in a general corpus, The inverse document frequency is used in the corpus of wind power. The slope of the life curve is obtained by statistically analyzing the word frequency changes of candidate words over the past few years, performing linear regression analysis on the word frequency changes, and extracting the slope of the regression line as the value of the indicator.

[0010] In some specific embodiments, the weighted fusion of three-dimensional feature indicators is performed as follows: the domain concentration (DR) value, the inverse document frequency difference (ΔIDF) and the slope of the life curve are standardized respectively, and then the standardized indicators are linearly weighted and summed according to preset weights to obtain the comprehensive domain score.

[0011] In some specific embodiments, in S3, words with a comprehensive domain score higher than the adaptive threshold θ are marked as pseudo-positive samples, and words with a comprehensive domain score lower than the adaptive threshold θ and that meet the high-frequency deactivation condition are marked as pseudo-negative samples.

[0012] In some specific embodiments, in S4, the structure of the character-level convolutional neural network model includes: an input layer, several one-dimensional convolutional layers, a global average pooling layer, and a fully connected output layer; the number of one-dimensional convolutional layers is 3, which are used to extract character-level local contextual features of candidate words.

[0013] In some specific embodiments, the preset iteration threshold in S4 is set to 0.85, and the preset stopping condition is that the number of iterations reaches 2 to 3 rounds.

[0014] According to a second aspect of the invention, a computer-readable storage medium is provided on which one or more computer programs are stored, which, when executed by a computer processor, implement the method described above.

[0015] According to a third aspect of the present invention, a zero-sample domain dictionary automatic construction system for wind power generation fault diagnosis is proposed, comprising: The data preprocessing and initial screening module is configured to receive raw text data in the field of wind power operation and maintenance. After preprocessing the raw text data, it extracts n-gram segments as initial candidate words, calculates the mutual information and left and right neighbor entropy of each initial candidate word, and retains high-quality candidate words based on preset statistical screening conditions. The multi-dimensional feature calculation module is configured to calculate three-dimensional feature indicators for each high-quality candidate word and perform weighted fusion to obtain the comprehensive domain score of the candidate word. The pseudo-label generation module is configured to use the comprehensive domain score distribution based on high-quality candidate words. It automatically calculates the optimal adaptive threshold to distinguish between domain words and non-domain words using the maximum inter-class variance method, and automatically classifies high-quality candidate words according to the adaptive threshold θ to construct a pseudo-label training set. The model self-training iteration module is configured to build a character-level convolutional neural network model and supervise the training of the model using a pseudo-label training set. The trained character-level convolutional neural network model is used to predict all high-quality candidate words. Words with prediction confidence higher than the preset iteration threshold are selected to expand the pseudo-label training set, and the expanded training set is used to train the model for the next round of iterations until the preset stopping condition is reached. The dictionary generation module is configured to output the top-ranked words in the final prediction confidence ranking, generating a dictionary for wind power fault diagnosis that includes word names, confidence scores, and part-of-speech tags.

[0016] In some specific embodiments, the system adopts an edge-cloud collaborative architecture: the data preprocessing and initial screening module and the multi-dimensional feature calculation module are deployed on the edge devices of the wind farm, and the edge devices are configured to upload only the extracted n-gram fragments and the calculated feature vectors to the cloud; the pseudo-label generation module, the model self-training iteration module and the dictionary generation module are deployed on the cloud server, and the cloud server is configured to receive the feature vectors and perform subsequent processing.

[0017] Compared with the prior art, the present invention has the following beneficial effects: This invention achieves fully automated dictionary construction under zero-sample conditions, completely solving the "cold start" problem. It abandons the reliance of traditional methods on expert-compiled dictionaries, manually labeled data, or initial seed dictionaries. By combining unsupervised statistical filtering with Otsu's adaptive threshold segmentation technique, it can automatically generate a high-quality dictionary directly from raw maintenance text in initial projects without any prior knowledge. This breaks the vicious cycle faced by traditional NLP systems in the wind power field: "no dictionary leads to difficulty in processing, and difficulty in processing leads to a lack of knowledge."

[0018] Unlike existing technologies based on static corpus analysis, this invention creatively introduces the time dimension feature of "life curve slope." By performing linear regression analysis on word frequencies, it can sensitively identify emerging terms that show an increasing trend with technological upgrades or the occurrence of failures. This makes the generated dictionary dynamically adaptable, automatically iterating along with wind turbine model updates and failure evolution.

[0019] To address the shortcomings of traditional statistical methods such as mutual information in distinguishing between "general high-frequency words" (such as "inspection" and "repair") and "domain-specific terms", this invention proposes a three-dimensional fusion scoring model consisting of "domain concentration (DR) + domain specificity (ΔIDF) + life curve slope". This model can accurately evaluate the domain attributes of words from multiple dimensions, effectively filter out general noise, and ensure the professionalism of the selected words.

[0020] This invention utilizes pseudo-labels automatically generated by the Otsu algorithm to train a character-level CNN model, and continuously optimizes the model performance through iterative self-training. This closed-loop mechanism enables the system to discover low-frequency or long-tail terms with inconspicuous statistical features, improving the F1 score of dictionary recognition by 8%–12% compared to purely statistical methods.

[0021] To address the limited network environment in wind farms, this invention designs a collaborative architecture of "edge computing statistical features + cloud-based deep learning." The edge side only needs to upload n-gram fragments and lightweight feature vectors, with data transmission volume less than 1% of the original data, significantly reducing network bandwidth requirements. This makes the solution highly suitable for low-cost deployment and rapid updates in remote wind farms. Attached Figure Description

[0022] The accompanying drawings are included to provide a further understanding of the embodiments and are incorporated in and constitute a part of this specification. The drawings illustrate embodiments and, together with the description, serve to explain the principles of the invention. Other embodiments and many anticipated advantages of the embodiments will be readily recognized as they become better understood through reference to the following detailed description. Other features, objects, and advantages of this application will become more apparent from reading the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is an exemplary system architecture diagram to which this application can be applied; Figure 2 This is a flowchart of a zero-sample domain dictionary automatic construction method for wind power generation fault diagnosis according to an embodiment of this application; Figure 3 This is a framework diagram of a zero-sample domain dictionary automatic construction system for wind power generation fault diagnosis, according to an embodiment of this application. Figure 4 This is a schematic diagram of the structure of a computer system used to implement the electronic device of the present application. Detailed Implementation

[0023] 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.

[0024] 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.

[0025] Figure 1An exemplary system architecture 100 for an automatic construction method of a zero-sample domain dictionary for wind power generation fault diagnosis, which can be applied according to embodiments of this application, is shown.

[0026] like Figure 1 As shown, system architecture 100 may include terminal devices 101, 102, and 103, a network 104, and a server 105. Network 104 serves as the medium for providing communication links between terminal devices 101, 102, and 103 and server 105. Network 104 may include various connection types, such as wired or wireless communication links, or fiber optic cables, etc.

[0027] Users can use terminal devices 101, 102, and 103 to interact with server 105 via network 104 to receive or send messages, etc. Various applications can be installed on terminal devices 101, 102, and 103, such as data processing applications, data visualization applications, and web browser applications.

[0028] Terminal devices 101, 102, and 103 can be either hardware or software. When terminal devices 101, 102, and 103 are hardware, they can be various electronic devices, including but not limited to smartphones, tablets, laptops, and desktop computers. When terminal devices 101, 102, and 103 are software, they can be installed in the electronic devices listed above. They can be implemented as multiple software programs or software modules (e.g., software programs or software modules used to provide distributed services) or as a single software program or software module. No specific limitations are imposed here.

[0029] Server 105 can be a server that provides various services, such as a background information processing server that supports terminal devices 101, 102, and 103. The background information processing server can process the acquired wind power generation operation and maintenance text data.

[0030] It should be noted that the method provided in this application embodiment can be executed by server 105 or by terminal devices 101, 102, and 103. The corresponding device is generally set in server 105 or can be set in terminal devices 101, 102, and 103.

[0031] It should be noted that a server can be either hardware or software. When the server is hardware, it can be implemented as a distributed server cluster consisting of multiple servers, or as a single server. When the server is software, it can be implemented as multiple software programs or software modules (such as software programs or software modules used to provide distributed services), or as a single software program or software module. No specific limitations are made here.

[0032] It should be understood that Figure 1The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0033] Figure 2 A flowchart illustrating an automatic zero-sample domain dictionary construction method for wind power generation fault diagnosis according to an embodiment of this application is shown. Figure 2 As shown, the method includes the following steps: S1: Receive raw text data in the field of wind power operation and maintenance, preprocess the raw text data and extract n-gram segments as initial candidate words, calculate the mutual information and left and right neighboring word entropy of each initial candidate word, and retain high-quality candidate words based on preset statistical screening conditions.

[0034] In a specific embodiment, the preprocessing of the original text data includes: receiving wind turbine maintenance work orders, technical notices, and operation and maintenance logs; unifying the text encoding to UTF-8 format to eliminate the impact of garbled characters; using the SimHash algorithm to deduplicate the text to avoid duplicate data causing deviations in word frequency statistics; filtering out text segments with a character length less than a preset lower limit (e.g., 10 characters) or greater than a preset upper limit (e.g., 512 characters), removing meaningless phrases or excessively long redundant paragraphs, and retaining sentences with dense effective information; and segmenting the text into independent sentence units based on punctuation marks such as periods, semicolons, and exclamation marks, which serve as the basic input for subsequent processing.

[0035] In a specific embodiment, the specific steps for extracting n-gram fragments and statistical filtering include: Sliding window extraction: The cleaned corpus is scanned with a sliding window to extract n-gram fragments of 2 to 6 characters in length as initial candidate words.

[0036] Minimum support filtering: Count the word frequency of each n-gram segment, set a minimum support threshold k, and retain only segments with a word frequency greater than k. Preferably, k = max( (Total number of characters in the corpus) is used to filter out extremely low-frequency, random string combinations.

[0037] Statistical Feature Calculation and Initial Screening: Calculate the cohesion and degrees of freedom for each candidate word. Cohesion: Use mutual information to measure the tightness of word-to-word combinations, filtering out "accidental co-occurrence" segments. Degrees of Freedom: Calculate the entropy values ​​of the left and right adjacent characters of a candidate word, taking the minimum of the left and right entropies as the degree of freedom index. This measures whether the word can be used independently and flexibly, thus filtering out pseudo-words that are "fixed collocations but cannot be further divided." Unsupervised Screening: Set dual thresholds based on statistical distribution (e.g., mean μ + standard deviation σ), retaining millions of high-quality candidate words whose mutual information and left and right entropies both meet the thresholds.

[0038] S2: For each high-quality candidate word obtained in step S1, calculate the three-dimensional feature index and perform weighted fusion to obtain the comprehensive domain score of the candidate word.

[0039] In a specific embodiment, the three-dimensional feature indicators include the Domain Concentration (DR) value, the Inverse Document Frequency Difference (ΔIDF), and the slope of the life curve. The formula for calculating the Domain Concentration (DR) value is as follows: ,in, The frequency of candidate words in the corpus of wind power-related terms. The term frequency of the candidate word in the general corpus; the formula for calculating the inverse document frequency difference ΔIDF is: ,in, Inverse document frequencies in a general corpus, The inverse document frequency is used in the corpus of wind power. The slope of the life curve is obtained by statistically analyzing the word frequency changes of candidate words over the past few years, performing linear regression analysis on the word frequency changes, and extracting the slope of the regression line as the value of the indicator.

[0040] In a specific embodiment, the weighted fusion of three-dimensional feature indicators is performed as follows: the domain concentration (DR) value, inverse document frequency difference (ΔIDF) and life curve slope are standardized (e.g., Z-score standardization) to eliminate dimensional differences. Then, the standardized indicators are linearly weighted and summed according to preset weights to obtain the comprehensive domain score.

[0041] S3: Based on the comprehensive domain score distribution of high-quality candidate words, the optimal adaptive threshold for distinguishing between domain words and non-domain words is automatically calculated using the maximum inter-class variance method. Based on the adaptive threshold θ, high-quality candidate words are automatically classified to construct a pseudo-label training set.

[0042] In a specific implementation, words with a comprehensive domain score higher than an adaptive threshold θ are marked as pseudo-positive samples, while words with a comprehensive domain score lower than the adaptive threshold θ and meeting the high-frequency deactivation condition are marked as pseudo-negative samples. This screening strategy ensures that negative samples are definite general words or noise, rather than potential long-tail domain words.

[0043] S4: Construct a character-level convolutional neural network model and supervise its training using a pseudo-label training set; use the trained character-level convolutional neural network model to predict all high-quality candidate words from step S1, select words with prediction confidence higher than the preset iteration threshold to expand the pseudo-label training set, and use the expanded training set to train the model for the next round of iterations until the preset stopping condition is reached.

[0044] In a specific embodiment, the structure of the character-level convolutional neural network model includes: an input layer, several one-dimensional convolutional layers, a global average pooling layer, and a fully connected output layer; the number of one-dimensional convolutional layers is 3, which are used to extract character-level local contextual features of candidate words.

[0045] In a specific embodiment, the preset iteration threshold is set to 0.85, and the preset stopping condition is that the number of iterations reaches 2 to 3 rounds. This prevents model drift while continuously improving recognition accuracy (F1 score can be improved by 8-12%).

[0046] S5: Output the top-ranked words in the final prediction confidence score, and generate a dictionary for wind power fault diagnosis that includes word names, confidence scores, and part-of-speech tags.

[0047] In a specific implementation, the final output dictionary format includes Top-K high-scoring terms. Each term not only contains the word itself but also metadata such as confidence score, part of speech (e.g., noun-component, verb-fault action) and the year of its first occurrence. This dictionary supports exporting to custom dictionary formats for common word segmentation tools such as jieba and pkuseg, allowing for direct integration into downstream applications.

[0048] Furthermore, this embodiment also supports an edge-cloud collaborative architecture: data cleaning, n-gram extraction, and statistical feature (mutual information, entropy) calculation are completed on the edge devices of the wind farm, and only the extracted n-gram list and its corresponding 6-dimensional feature vector (the data volume is less than 1% of the original text) are uploaded to the cloud; the cloud server is responsible for performing CNN model training, threshold calculation, and final dictionary generation. This architecture significantly reduces network transmission overhead and enables low-cost engineering implementation.

[0049] Figure 3 This paper illustrates a framework diagram of a zero-sample domain dictionary automatic construction system for wind power generation fault diagnosis, based on an embodiment of this application. Figure 3As shown, the system mainly includes a data preprocessing and initial screening module 301, a multi-dimensional feature calculation module 302, a pseudo-label generation module 303, a model self-training iteration module 304, and a dictionary generation module 305. Specifically, the data preprocessing and initial screening module 301 is configured to receive raw text data from the wind power generation operation and maintenance field, preprocess the raw text data to extract n-gram segments as initial candidate words, calculate the mutual information and left and right neighbor entropies of each initial candidate word, and retain high-quality candidate words based on preset statistical screening conditions. The multi-dimensional feature calculation module 302 is configured to calculate three-dimensional feature indicators for each high-quality candidate word and perform weighted fusion to obtain the comprehensive domain score of the candidate word. The pseudo-label generation module 303 is configured to automatically calculate the optimal adaptive threshold for distinguishing domain words from non-domain words based on the comprehensive domain score distribution of high-quality candidate words, using the maximum inter-class variance method, and automatically classify high-quality candidate words according to the adaptive threshold θ to construct a pseudo-label training set. The model self-training iteration module 304 is configured to construct a character-level convolutional neural network model and use the pseudo-label training set for supervised training of the model. The trained character-level convolutional neural network model is used to predict all high-quality candidate words. Words with prediction confidence higher than the preset iteration threshold are selected and expanded into the pseudo-label training set. The expanded training set is then used to train the model for the next round of iterations until the preset stopping condition is reached. The dictionary generation module 305 is configured to output the words with the highest prediction confidence, generating a dictionary for wind power generation fault diagnosis that includes word names, confidence levels, and part-of-speech tags.

[0050] This application proposes a zero-shot domain dictionary automatic construction method and system for wind power generation fault diagnosis, aiming to address the pain points of existing technologies, such as heavy reliance on expert manual maintenance, difficulties in cold start, and difficulty in capturing emerging fault terms. It constructs a three-dimensional domain-specific evolutionary feature model including the "life curve slope," capable of dynamically identifying technical terms that evolve over time. Simultaneously, it creatively utilizes the Otsu's method to automatically generate pseudo-labels, combined with a self-training iterative closed loop of a character-level CNN model, achieving fully automated construction without any initial dictionary or manual annotation. Furthermore, the adoption of an edge-cloud collaborative architecture significantly reduces data transmission costs.

[0051] The following is for reference. Figure 4 It shows a schematic diagram of the structure of a computer system suitable for implementing the electronic device of the present application. Figure 4 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.

[0052] like Figure 4As shown, the computer system includes a central processing unit (CPU) 401, which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) 402 or programs loaded from storage section 408 into random access memory (RAM) 403. RAM 403 also stores various programs and data required for the operation of system 400. CPU 401, ROM 402, and RAM 403 are interconnected via bus 404. Input / output (I / O) interface 405 is also connected to bus 404.

[0053] The following components are connected to I / O interface 405: an input section 406 including a keyboard, mouse, etc.; an output section 407 including a liquid crystal display (LCD) and speakers, etc.; a storage section 408 including a hard disk, etc.; and a communication section 409 including a network interface card such as a LAN card and a modem, etc. The communication section 409 performs communication processing via a network such as the Internet. Drive 410 is also connected to I / O interface 405 as needed. Removable media 411, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 410 as needed so that computer programs read from them can be installed into storage section 408 as needed.

[0054] 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 409, and / or installed from removable medium 411. When the computer program is executed by central processing unit (CPU) 401, 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 thereof. 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.

[0055] 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).

[0056] 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.

[0057] The modules described in the embodiments of this application can be implemented in software or in hardware.

[0058] 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 aforementioned one or more programs are executed by the electronic device, the electronic device causes the following: receiving raw text data from the wind power generation operation and maintenance field; preprocessing the raw text data and extracting n-gram segments as initial candidate words; calculating the mutual information and left and right neighbor entropies of each initial candidate word; retaining high-quality candidate words based on preset statistical screening conditions; calculating three-dimensional feature indicators and performing weighted fusion for each high-quality candidate word obtained in step S1 to obtain the comprehensive domain score of the candidate word; automatically calculating the optimal adaptive threshold for distinguishing domain words from non-domain words using the maximum inter-class variance method based on the comprehensive domain score distribution of the high-quality candidate words; automatically classifying the high-quality candidate words according to the adaptive threshold θ; constructing a pseudo-label training set; constructing a character-level convolutional neural network model; and supervising the model training using the pseudo-label training set. The trained character-level convolutional neural network model is used to predict all high-quality candidate words in step S1. Words with prediction confidence higher than the preset iteration threshold are selected and expanded into the pseudo-label training set. The expanded training set is used to train the model in the next round of iterations until the preset stopping condition is reached. The words with the highest prediction confidence are output, and a dictionary for wind power fault diagnosis containing word names, confidence levels and part-of-speech tags is generated.

[0059] 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 features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A method for automatic construction of a zero-sample domain dictionary for fault diagnosis in wind power generation, characterized in that, include: S1: Receive raw text data in the field of wind power operation and maintenance, preprocess the raw text data and extract n-gram segments as initial candidate words, calculate the mutual information and left and right neighboring word entropy of each initial candidate word, and retain high-quality candidate words based on preset statistical screening conditions. S2: For each high-quality candidate word obtained in step S1, calculate the three-dimensional feature index and perform weighted fusion to obtain the comprehensive domain score of the candidate word; S3: Based on the comprehensive domain score distribution of the high-quality candidate words, the optimal adaptive threshold for distinguishing between domain words and non-domain words is automatically calculated using the maximum inter-class variance method. The high-quality candidate words are automatically classified according to the adaptive threshold θ to construct a pseudo-label training set. S4: Construct a character-level convolutional neural network model and supervise the training of the model using the pseudo-label training set; use the trained character-level convolutional neural network model to predict all the high-quality candidate words in step S1, select words with prediction confidence higher than the preset iteration threshold to expand the pseudo-label training set, and use the expanded training set to conduct the next round of iteration training of the model until the preset stopping condition is reached. S5: Output the top-ranked words in the final prediction confidence score, and generate a dictionary for wind power fault diagnosis that includes word names, confidence scores, and part-of-speech tags.

2. The method for automatic construction of a zero-sample domain dictionary for wind power generation fault diagnosis according to claim 1, characterized in that, The preprocessing of the original text data in S1 specifically includes: receiving wind turbine maintenance work orders, technical notices, and operation and maintenance logs, and unifying the text encoding to UTF-8 format; using the SimHash algorithm to deduplicate the text; filtering out text segments with character lengths less than a preset lower limit or greater than a preset upper limit; and segmenting the text into independent sentences based on punctuation marks.

3. The method for automatic construction of a zero-sample domain dictionary for wind power generation fault diagnosis according to claim 1, characterized in that, The three-dimensional feature indicators in S2 include the Domain Concentration (DR) value, the Inverse Document Frequency Difference (ΔIDF), and the slope of the life curve. The formula for calculating the Domain Concentration (DR) value is as follows: ,in, The frequency of candidate words in the corpus of wind power-related terms. The term frequency of the candidate word in the general corpus; the formula for calculating the inverse document frequency difference ΔIDF is: ,in, Inverse document frequencies in a general corpus, The inverse document frequency is the frequency of a corpus in the wind power field. The slope of the life curve is obtained by: statistically analyzing the word frequency change data of candidate words in recent years, performing linear regression analysis on the word frequency change data, and extracting the slope of the regression line as the value of the index.

4. The method for automatic construction of a zero-sample domain dictionary for wind power generation fault diagnosis according to claim 3, characterized in that, The specific method for weighted fusion of the three-dimensional feature indicators is as follows: the domain concentration (DR) value, inverse document frequency difference (ΔIDF) and life curve slope are standardized respectively, and then the standardized indicators are linearly weighted and summed according to preset weights to obtain the comprehensive domain score.

5. The method for automatic construction of a zero-sample domain dictionary for wind power generation fault diagnosis according to claim 1, characterized in that, In S3, words with a comprehensive domain score higher than the adaptive threshold θ are marked as pseudo-positive samples, and words with a comprehensive domain score lower than the adaptive threshold θ and that meet the high-frequency deactivation condition are marked as pseudo-negative samples.

6. The method for automatic construction of a zero-sample domain dictionary for wind power generation fault diagnosis according to claim 1, characterized in that, In S4, the structure of the character-level convolutional neural network model includes: an input layer, several one-dimensional convolutional layers, a global average pooling layer, and a fully connected output layer; the one-dimensional convolutional layers have three layers and are used to extract character-level local contextual features of candidate words.

7. The method for automatic construction of a zero-sample domain dictionary for wind power generation fault diagnosis according to claim 1, characterized in that, The preset iteration threshold in S4 is set to 0.85, and the preset stopping condition is that the number of iterations reaches 2 to 3 rounds.

8. A computer-readable storage medium having one or more computer programs stored thereon, characterized in that, When the one or more computer programs are executed by a computer processor, they perform the method according to any one of claims 1-7.

9. A zero-sample domain dictionary automatic construction system for wind power generation fault diagnosis, characterized in that, include: The data preprocessing and initial screening module is configured to receive raw text data in the field of wind power operation and maintenance, preprocess the raw text data and extract n-gram segments as initial candidate words, calculate the mutual information and left and right neighboring word entropy of each initial candidate word, and retain high-quality candidate words based on preset statistical screening conditions. The multi-dimensional feature calculation module is configured to calculate three-dimensional feature indicators for each high-quality candidate word and perform weighted fusion to obtain the comprehensive domain score of the candidate word. The pseudo-label generation module is configured to use the comprehensive domain score distribution based on the high-quality candidate words, automatically calculate the optimal adaptive threshold for distinguishing domain words from non-domain words using the maximum inter-class variance method, automatically classify the high-quality candidate words according to the adaptive threshold θ, and construct a pseudo-label training set. The model self-training iteration module is configured to build a character-level convolutional neural network model and supervise the training of the model using the pseudo-label training set. The trained character-level convolutional neural network model is used to predict all high-quality candidate words, and words with prediction confidence higher than a preset iteration threshold are selected to expand the pseudo-label training set. The expanded training set is then used to train the model for the next round of iterations until a preset stopping condition is reached. The dictionary generation module is configured to output the top-ranked words in the final prediction confidence ranking, generating a dictionary for wind power fault diagnosis that includes word names, confidence scores, and part-of-speech tags.

10. The zero-sample domain dictionary automatic construction system for wind power generation fault diagnosis according to claim 9, characterized in that, The system adopts an edge-cloud collaborative architecture: the data preprocessing and initial screening module and the multi-dimensional feature calculation module are deployed on the edge devices of the wind farm. The edge devices are configured to upload only the extracted n-gram fragments and the calculated feature vectors to the cloud. The pseudo-label generation module, the model self-training iteration module and the dictionary generation module are deployed on the cloud server. The cloud server is configured to receive the feature vectors and perform subsequent processing.

Images