An early fault detection system for wind turbine generator sets

Abstract

The present invention relates to the technical field of wind turbine generator set detection, and specifically to a wind turbine generator set early fault detection system, comprising: a preliminary detection unit, for obtaining current operating data of the wind turbine generator set, and inputting the current operating data into a pre-trained fault feature recognition model to obtain the component name and preliminary fault feature of each first fault component; a text generation unit, for obtaining the local state of the unit corresponding to the current operating data; wherein the local state of the unit includes first time series feature data corresponding to each first fault component. By inputting the current operating data of the wind turbine generator set into a pre-trained fault feature recognition model, the present invention can accurately identify the preliminary fault features of each component. This approach can help monitor the operating status of the unit in real time, discover potential fault risks in advance, and thus avoid damage or shutdown of the equipment.

Description

Technical Field

[0001] The present invention relates to the technical field of wind turbine generator set detection, and in particular to an early fault detection system for a wind turbine generator set. Background Art

[0002] Traditional systems typically rely on manual inspections, regular inspections, and empirical judgment. This approach is not only limited by the accuracy and reliability of manual operations, but is also susceptible to human negligence or misjudgment, making it impossible to achieve real-time monitoring and comprehensive coverage, resulting in the failure to detect potential faults in a timely manner. Traditional systems often lack real-time data monitoring and automatic fault diagnosis capabilities, and are unable to provide timely warnings at the early stages of faults. Fault detection is usually based on post-analysis, missing the optimal maintenance window. This delayed response can cause equipment damage or shutdown, increasing maintenance and downtime. Traditional systems often rely on simple fault diagnosis standards and manual analysis, unable to fully explore the time series characteristic data of various components of wind turbines, and unable to perform large-scale historical data comparison and analysis. Traditional systems usually do not have intelligent analysis models, lack automatically updated databases, and lack self-learning capabilities. As equipment and fault types change, databases easily become outdated, resulting in fault diagnosis no longer adapting to new equipment conditions and fault characteristics, thereby affecting their long-term effectiveness. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to overcome the shortcomings of the above-mentioned prior art and provide a wind turbine generator set early fault detection system.

[0004] The technical solution adopted to solve the above technical problems is: a wind turbine generator set early fault detection system, comprising:

[0005] a preliminary detection unit, configured to obtain current operating data of the wind turbine generator set, and input the current operating data into a pre-trained fault feature recognition model to obtain a component name and preliminary fault feature of each first faulty component;

[0006] a text generation unit configured to obtain a local state of the unit corresponding to the current operating data; wherein the local state of the unit includes first time series feature data corresponding to each first faulty component; and input each component name and preliminary fault feature into a large language model to obtain a fault location text and a fault attribute text for each first faulty component;

[0007] A fault storage unit, configured to obtain a global fault database of the wind turbine generator set; wherein the global fault database includes second time series feature data corresponding to a plurality of second fault components, historical fault location texts, historical fault attribute texts, historical component names, and historical semantic feature vectors;

[0008] a storage and updating unit, configured to match the first faulty component with the global fault database to obtain a second faulty component that matches the first faulty component, and use the second faulty component as a first target faulty component; and perform feature fusion on the first faulty component and the first target faulty component to update the global fault database;

[0009] An early detection unit is used to locate each component to be inspected in the wind turbine generator set; obtain a retrieval feature of the component to be inspected; and match the retrieval feature with the updated global fault database to obtain a second target fault component that matches the component to be inspected.

[0010] Preferably, performing feature fusion on the first faulty component and the first target faulty component to update the global fault database includes:

[0011] fusing the first time series feature data corresponding to the first faulty component with the second time series feature data corresponding to the first target faulty component;

[0012] Merging the fault location text, fault attribute text, and component name corresponding to the first faulty component into the historical fault location text, historical fault attribute text, and historical component name corresponding to the first target faulty component;

[0013] The preliminary fault features and component names of each first fault component are input into a pre-trained semantic extraction model to obtain a semantic feature vector corresponding to each first fault component.

[0014] Preferably, performing feature fusion on the first faulty component and the first target faulty component to update the global fault database further includes:

[0015] Searching for a first target fault component that matches each first fault component among a plurality of second fault components included in the global fault database, and fusing semantic feature vectors corresponding to the first target fault components into a historical semantic feature vector corresponding to the first target fault component;

[0016] The second time series feature data, historical fault location text, historical fault attribute text, historical component name and historical semantic feature vector corresponding to the first target fault component are added to the global fault database to update the global fault database.

[0017] Preferably, the fault location text, fault attribute text, and component name corresponding to the first fault component are respectively integrated into the historical fault location text, historical fault attribute text, and historical component name corresponding to the first target fault component, including:

[0018] Searching for a standard component name that matches the component name of the first fault component in a preset wind turbine generator set standard fault table;

[0019] Merging the standard component name of the first faulty component into the historical component name corresponding to the first target faulty component;

[0020] The fault location text and the fault attribute text of the first fault component are added to the historical fault location text and the historical fault attribute text corresponding to the first target fault component.

[0021] Preferably, performing feature fusion on the first faulty component and the first target faulty component to update the global fault database further includes:

[0022] When the first early fault detection time of the historical fault location text corresponding to the second fault component is before the preset maintenance cycle, and the number of early fault detections of the historical fault attribute text corresponding to the second fault component is less than the first preset threshold, the historical fault location text corresponding to the second fault component is deleted from the updated global fault database.

[0023] Preferably, matching the first faulty component with the global fault database to obtain a second faulty component that matches the first faulty component, and using the second faulty component as a first target faulty component, comprises:

[0024] Determine a timing correlation and a first feature similarity between each second fault component and the first fault component in the global fault database;

[0025] determining a target correlation based on a timing correlation and a first feature similarity between the second faulty component and the first faulty component;

[0026] A second fault component whose target correlation with the first fault component is greater than a second preset threshold among the second fault components in the global fault database is determined as a first target fault component matching the first fault component.

[0027] Preferably, obtaining the local state of the unit corresponding to the current operating data includes:

[0028] Obtaining vibration time series data and temperature time series data corresponding to the current operating data;

[0029] generating a time-frequency characteristic graph corresponding to the current operation data according to the vibration time series data and the temperature time series data;

[0030] The local state of the unit is generated according to the time-frequency characteristic diagram, and first time series characteristic data corresponding to the first faulty component is generated in the local state.

[0031] Preferably, the retrieval feature includes at least one of time series feature data, text feature and keyword feature of component name; wherein, the text feature includes fault location text feature and fault attribute text feature.

[0032] Preferably, matching the search feature with the updated global fault database to obtain a second target faulty component that matches the component to be inspected includes:

[0033] Determine a timing matching degree between the timing characteristic data and each second fault component in the updated global fault database;

[0034] Determining a semantic match between the text feature and each second fault component in the updated global fault database;

[0035] Determine a second target faulty component in the updated global fault database that matches the search feature according to the timing matching degree and the semantic matching degree corresponding to the second faulty component.

[0036] Preferably, matching the search feature with the updated global fault database to obtain a second target faulty component that matches the component to be inspected further includes:

[0037] When the number of keywords included in the search feature is less than or equal to a third preset threshold, determining a keyword matching degree between the keyword feature and each second fault component in the updated global fault database;

[0038] When the number of keywords included in the search feature is less than or equal to a third preset threshold, a second target faulty component matching the search feature in the updated global fault database is determined according to the timing matching degree and keyword matching degree corresponding to the second faulty component.

[0039] The beneficial effects of the present invention are as follows: (1) By inputting the current operating data of the wind turbine generator set into a pre-trained fault feature recognition model, the system can accurately identify the preliminary fault features of each component. This approach can help monitor the operating status of the unit in real time and discover potential fault risks in advance, thereby avoiding equipment damage or shutdown; and by collecting the time series feature data of each faulty component of the wind turbine generator set and comparing it with historical data, the faulty component can be located and analyzed. This method helps to more accurately identify the time, frequency and possible causes of the fault, thereby improving the timeliness and accuracy of fault diagnosis; (2) By using a large language model to generate fault location and fault attribute text, the system can further describe the fault in detail, which is This text generation method not only improves the interpretability of diagnosis, but also provides technicians with more intuitive fault information, facilitating faster decision-making. In the detection process, by matching the newly identified fault features with the data in the global fault database and updating the database through feature fusion, this dynamic update mechanism ensures that the fault database can maintain high sensitivity to future potential faults while continuously acquiring new data, avoiding diagnostic errors caused by outdated data. (3) The present invention can monitor each component to be inspected in real time, obtain its retrieval features and match them with the updated fault database. This function can detect hidden danger components in advance and ensure that preventive maintenance measures are taken before a fault occurs, thereby extending the service life of the equipment, reducing maintenance costs, and ensuring efficient operation of the unit. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a schematic diagram of the system architecture of the overall system in an embodiment of the present invention.

[0041] Figure numerals: 1. Preliminary detection unit; 2. Text generation unit; 3. Fault storage unit; 4. Storage update unit; 5. Early detection unit. DETAILED DESCRIPTION

[0042] Example 1, as Figure 1 As shown, the present invention proposes a wind turbine generator set early fault detection system, comprising:

[0043] A preliminary detection unit 1 is used to obtain current operating data of the wind turbine generator set and input the current operating data into a pre-trained fault feature recognition model to obtain the component name and preliminary fault feature of each first fault component;

[0044] Text generation unit 2 is configured to obtain a local state of the unit corresponding to the current operating data; wherein the local state of the unit includes first time series feature data corresponding to each first faulty component; and input each component name and preliminary fault feature into the large language model to obtain a fault location text and a fault attribute text for each first faulty component.

[0045] The fault storage unit 3 is used to obtain a global fault database of the wind turbine generator set; wherein the global fault database includes second time series feature data corresponding to multiple second fault components, historical fault location texts, historical fault attribute texts, historical component names, and historical semantic feature vectors;

[0046] The storage and updating unit 4 is configured to match the first faulty component with the global fault database to obtain a second faulty component that matches the first faulty component, and use the second faulty component as the first target faulty component; perform feature fusion on the first faulty component and the first target faulty component to update the global fault database;

[0047] The early detection unit 5 is used to locate each component to be inspected in the wind turbine generator set; obtain the retrieval feature of the component to be inspected; and match the retrieval feature with the updated global fault database to obtain a second target fault component that matches the component to be inspected.

[0048] In the present invention, the current operating data of the wind turbine generator set is obtained; these current operating data are input into a pre-trained fault feature recognition model, which is trained to be able to identify the preliminary fault features of different components; the model outputs the component name and preliminary fault features related to each first fault component (i.e., the component currently faulty); sensors (such as vibration sensors, temperature sensors, etc.) are used to obtain the operating data of the unit in real time; the fault feature recognition model is trained using a machine learning algorithm (such as random forest, SVM, deep neural network, etc.), and the model should be able to perform fault diagnosis based on historical data and features; a large amount of labeled fault data (such as historical cases of certain known faults) is used to train the model so that it can make accurate fault predictions for real-time data; the local state of the unit corresponding to the current operating data is obtained, which usually involves the time series feature data of specific components (such as the operating curves, pressure, temperature, etc. of each component that change with time); a time series analysis method (such as dynamic time warping, Fourier transform or wavelet transform, etc.) is used to extract the time series features of each component; the component name and preliminary fault features of each component are respectively input into a large language model (such as GPT, BERT, etc.), which can Based on this information, fault location text and fault attribute text are generated. Based on historical fault data from wind turbines, the large language model is fine-tuned to accurately generate textual information related to component faults (e.g., "The generator fault of unit X occurred at the rotor blade connection, and the fault was overheating"). Through continuous monitoring and fault logging, fault information for each component is accumulated to form a global database. This database should not only contain traditional time series data but also natural language descriptions and semantic feature vectors to facilitate more in-depth analysis. A pre-trained semantic model (such as Word2Vec or BERT) is used to encode historical data and convert text into semantic feature vectors. The first faulty component (the current faulty component) is matched with historical faulty components in the global fault database to identify the second faulty component that best matches it and use it as the first target faulty component. Feature fusion (e.g., time series data, textual data, semantic vectors, etc.) is performed on these two faulty components to update the global fault database. The location of the component to be inspected is determined using sensors and positioning technologies (such as GPS and laser scanning). The features of the component to be inspected are compared with historical fault data in the database to output the most likely fault type and component.

[0049] In an optional embodiment, performing feature fusion on the first faulty component and the first target faulty component to update the global fault database includes:

[0050] fusing the first time series feature data corresponding to the first faulty component with the second time series feature data corresponding to the first target faulty component;

[0051] Merging the fault location text, fault attribute text, and component name corresponding to the first fault component into the historical fault location text, historical fault attribute text, and historical component name corresponding to the first target fault component, respectively;

[0052] The preliminary fault features and component names of each first fault component are input into a pre-trained semantic extraction model to obtain a semantic feature vector corresponding to each first fault component.

[0053] It should be noted that the first faulty component is the component that is currently faulty and has time series feature data associated with it (such as temperature, vibration, and other data that change over time); the first target faulty component is the most similar component obtained by matching with historical faulty components in the global fault database; here, the time series feature data of the first faulty component is fused with the historical time series feature data of the first target faulty component. Through the combination of time series data, the temporal variation trend of the fault feature can be fully understood, which helps to enhance the diagnostic accuracy of the current fault; the fault location text and the fault attribute text are natural language descriptions that describe the specific characteristics of the fault (such as "the bearing of the generator has an overheating fault"); the component name refers to the name of the faulty component (such as "generator" and "wind rotor blade", etc.); in this step, the fault information of the first faulty component (such as fault location, The fault attributes and component names are then text-fused with the historical fault information of the first target fault component. This fusion allows the specific fault condition of the current component to be compared and combined with relevant information from historical cases, providing the system with a more comprehensive fault description. Preliminary fault features are the fault modes (such as "overheating" and "abnormal vibration") initially identified by the machine learning model. Component names refer to the specific components that have failed. Semantic extraction models, typically deep learning-based natural language processing models (such as BERT and GPT), convert text data into semantic feature vectors that better represent the semantic information in fault features and component names. In this step, the fault features and component names of each first fault component are input into the trained semantic extraction model to obtain semantic feature vectors for the fault component. These vectors are abstract representations of the fault mode and component name and can be used for further analysis and matching.

[0054] In an optional embodiment, performing feature fusion on the first faulty component and the first target faulty component to update the global fault database further includes:

[0055] Searching for a first target fault component that matches each first fault component among multiple second fault components included in the global fault database, and fusing the semantic feature vector corresponding to the first target fault component into the historical semantic feature vector corresponding to the first target fault component;

[0056] The second time series feature data, historical fault location text, historical fault attribute text, historical component name and historical semantic feature vector corresponding to the first target fault component are added to the global fault database to update the global fault database.

[0057] It should be noted that each faulty component in the database has a semantic feature vector, which represents the semantic information of the faulty component and is usually extracted through a deep learning model (such as BERT). The semantic feature vector of the first target faulty component is fused with the historically related semantic feature vectors of the component. This fusion helps capture the association between new fault modes and historical modes, and improves the system's understanding of the causes and types of faults.

[0058] In an optional embodiment, the fault location text, fault attribute text, and component name corresponding to the first fault component are respectively merged into the historical fault location text, historical fault attribute text, and historical component name corresponding to the first target fault component, including:

[0059] Searching for a standard component name that matches the component name of the first fault component in a preset wind turbine generator set standard fault table;

[0060] Merging the standard component name of the first faulty component into the historical component name corresponding to the first target faulty component;

[0061] The fault location text and the fault attribute text of the first fault component are added to the historical fault location text and the historical fault attribute text corresponding to the first target fault component.

[0062] It should be noted that in the system, there is a standard fault table containing various types of faults in advance; this table lists the standard components of all wind turbines and their corresponding fault information; the first fault component is the component that currently fails; the system will search for the standard component name that matches the component name in the standard fault table, in order to establish a connection between the current fault component and the standard component; the standard component name of the current first fault component is integrated into the historical component name of the first target fault component, in order to combine the new fault information with the existing fault data to improve the accuracy and reliability of fault prediction; the fault location text of the first fault component refers to the description of the specific location when the component fails (such as the location of the unit, the fault location of a specific component, etc.); the fault attribute text describes the type, severity and other characteristics of the fault; adding these new fault locations and attribute information to the historical fault data corresponding to the first target fault component can help the system better analyze the fault mode; the second fault component refers to other fault components in the database; the first early fault detection time of the historical fault location text refers to the time when the second fault component first had an early fault in the past records.

[0063] In an optional embodiment, performing feature fusion on the first faulty component and the first target faulty component to update the global fault database further includes:

[0064] When the first early fault detection time of the historical fault location text corresponding to the second fault component is before the preset maintenance cycle, and the number of early fault detections of the historical fault attribute text corresponding to the second fault component is less than the first preset threshold, the historical fault location text corresponding to the second fault component will be deleted from the updated global fault database.

[0065] It should be noted that the preset maintenance cycle is a fixed time range, usually referring to the maintenance cycle of the equipment (for example, maintenance is performed every 6 months); the number of early fault detections in the historical fault attribute text refers to the number of fault events detected within the cycle, which is usually counted by the fault warning system; the first preset threshold is a specific value used to determine whether the historical records need to be deleted; when the number of early fault detections is less than the threshold, it means that the failure of the faulty component may not have a significant impact, or the failure mode is not frequent enough, so its historical records can be deleted from the database.

[0066] In an optional embodiment, matching the first faulty component with a global fault database to obtain a second faulty component that matches the first faulty component, and using the second faulty component as a first target faulty component includes:

[0067] Determine a time sequence correlation and a first feature similarity between each second fault component and the first fault component in the global fault database;

[0068] determining a target correlation based on a timing correlation and a first feature similarity between the second faulty component and the first faulty component;

[0069] A second fault component whose target correlation with the first fault component in the global fault database is greater than a second preset threshold is determined as a first target fault component matching the first fault component.

[0070] It should be noted that in the global fault database, each faulty component has a historical record and related information. The goal is to compare and match the current first faulty component with other components in the database to find a second faulty component that may have a similar failure mode. The temporal correlation measures the similarity in the time series of the fault occurrence between the second faulty component and the first faulty component. That is, whether the failure times of the two components are close, or whether the failure modes show similar trends over time. For example, if the failures of two components occur in the same time period or show similar patterns in the time window of the failure, then their temporal correlation is high. The first feature similarity refers to the similarity of the second faulty component based on some features of the first faulty component (such as fault type, fault location, fault nature, etc.). The "features" here may include the working conditions of the equipment, the impact of the fault, the symptoms manifested, etc. If the characteristics of the second faulty component are similar to those of the first faulty component, the similarity between them is high; the target correlation is a comprehensive indicator used to evaluate the overall matching degree between the second faulty component and the first faulty component; it is usually a weighted calculation result based on the timing correlation and feature similarity: for example, the timing correlation and feature similarity can be calculated by weighted average or other algorithms to obtain a comprehensive target correlation score. The higher the score, the higher the similarity between the two components in time and features, and the higher the matching degree; the second preset threshold is a set value used to determine which second faulty components have a high enough matching degree with the first faulty component and are worthy of being considered as a potential target faulty component.

[0071] In an optional embodiment, obtaining the local state of the unit corresponding to the current operating data includes:

[0072] Obtain the vibration time series data and temperature time series data corresponding to the current operating data;

[0073] Generate a time-frequency characteristic graph corresponding to the current operating data according to the vibration time series data and the temperature time series data;

[0074] A local state of the unit is generated according to the time-frequency characteristic diagram, and first time series characteristic data corresponding to the first faulty component is generated in the local state.

[0075] It should be noted that vibration time series data represents the change of vibration intensity of the machine over time during operation, which can help to judge the wear, failure or malfunction of mechanical parts; temperature time series data records the temperature change of the machine at different time points. Temperature anomalies are usually a harbinger of mechanical failure, equipment overload and other problems; time-frequency analysis refers to analyzing the signal in both time and frequency dimensions; traditional time domain analysis cannot provide information on frequency changes, and frequency domain analysis cannot reflect the change of the signal over time, so time-frequency analysis can combine the advantages of both; commonly used time-frequency analysis methods include short-time Fourier transform (STFT), wavelet transform, etc.; by performing time-frequency conversion on vibration time series data and temperature time series data, the information of the signal at different time and frequency can be extracted, thereby generating a time-frequency feature map, which can display possible frequency fluctuations, vibration intensity changes and other characteristics during machine operation; based on the time-frequency feature map, the status of the machine or equipment is analyzed, and the time-frequency characteristics The graph provides the frequency components and time dynamic characteristics of vibration and temperature signals, through which the health status or working status of the machine can be inferred; the local state refers to the health status of a specific part of the equipment (such as a specific component or subsystem), which is usually represented by time-frequency characteristics; it reflects whether the component is in normal working condition or whether there is an abnormality; under the known local state of the unit, the relationship between the local state and the first faulty component (i.e., the component that needs to be diagnosed) is analyzed; for example, the failure mode of a component may be pointed out through characteristics such as vibration signals and temperature changes; the first time series feature data refers to the specific time series features related to the first faulty component extracted through vibration, temperature and other data; these feature data reflect the abnormal performance of the component during operation, such as changes in vibration patterns, abnormal temperature fluctuations, etc.; the signal changes in the time-frequency feature graph can be used to determine whether the machine is working normally and to evaluate whether there are potential faults.

[0076] In an optional embodiment, the retrieval feature includes at least one of time series feature data, text feature and keyword feature of component name; wherein the text feature includes fault location text feature and fault attribute text feature.

[0077] In an optional embodiment, matching the search feature with the updated global fault database to obtain a second target faulty component that matches the component to be inspected includes:

[0078] Determine a timing matching degree between the timing characteristic data and each second fault component in the updated global fault database;

[0079] Determining a semantic match between the text feature and each second fault component in the updated global fault database;

[0080] A second target faulty component that matches the search feature in the updated global fault database is determined according to the timing matching degree and the semantic matching degree corresponding to the second faulty component.

[0081] It should be noted that temporal matching refers to how to measure the similarity between them by comparing the temporal feature data of the retrieval feature with the temporal feature data of each faulty component in the global fault database. Generally speaking, methods such as Euclidean distance and dynamic time warping (DTW) are used to measure the similarity of these temporal data. A high temporal matching degree indicates that the dynamic changes of the two data sets are very similar, which may represent similar fault modes. Semantic matching measures their semantic similarity by comparing these text features with the descriptive information in the global fault database. Semantic matching is usually achieved through natural language processing (NLP) technology, such as using word vectors or BERT models to quantify the similarity of texts. Components with similar text descriptions may have similar fault types. By combining temporal matching and semantic matching, the components in the database that are similar to the retrieval features are comprehensively evaluated. The higher the matching degree, the more similar the fault mode, cause, etc. are to the retrieval component, thereby determining the second target fault component that matches the retrieval feature.

[0082] In an optional embodiment, matching the search feature with the updated global fault database to obtain a second target faulty component that matches the component to be inspected further includes:

[0083] When the number of keywords included in the search feature is less than or equal to a third preset threshold, determining a keyword matching degree between the keyword feature and each second fault component in the updated global fault database;

[0084] When the number of keywords included in the search feature is less than or equal to the third preset threshold, a second target faulty component matching the search feature in the updated global fault database is determined according to the timing matching degree and the keyword matching degree corresponding to the second faulty component.

[0085] It should be noted that keyword features refer to keywords extracted from fault retrieval features, which may include component name, fault type, fault phenomenon and other information; keyword matching is performed when the number of keywords in the retrieval feature is less than or equal to the third preset threshold; this means that if the retrieval feature contains fewer keywords, these keywords can be matched with the fault components in the database to determine the similarity between them; the third preset threshold is used to determine whether the number of keywords is small enough; if so, the system will use the keyword matching method to match the fault component; if the number of keyword features is less than or equal to the third preset threshold, the system will use a combination of timing matching and keyword matching to determine the target fault component; in this way, even if there are only a small number of keywords in the retrieval feature, suitable fault components can still be found by matching timing features and keywords.

[0086] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A wind turbine generator set early fault detection system, characterized in that: include: a preliminary detection unit, configured to obtain current operating data of the wind turbine generator set, and input the current operating data into a pre-trained fault feature recognition model to obtain a component name and preliminary fault feature of each first faulty component; a text generation unit configured to obtain a local state of the unit corresponding to the current operating data; wherein the local state of the unit includes first time series feature data corresponding to each first faulty component; and input each component name and preliminary fault feature into a large language model to obtain a fault location text and a fault attribute text for each first faulty component; A fault storage unit, configured to obtain a global fault database of the wind turbine generator set; wherein the global fault database includes second time series feature data corresponding to a plurality of second fault components, historical fault location texts, historical fault attribute texts, historical component names, and historical semantic feature vectors; a storage and updating unit, configured to match the first faulty component with the global fault database to obtain a second faulty component that matches the first faulty component, and use the second faulty component as a first target faulty component; and perform feature fusion on the first faulty component and the first target faulty component to update the global fault database; The early detection unit is used to locate each component to be inspected in the wind turbine generator set; obtain the retrieval characteristics of the component to be inspected; and match the retrieval characteristics with the updated global fault database to obtain a second target fault component that matches the component to be inspected.

2. A wind turbine generator set early fault detection system according to claim 1, characterized in that: Performing feature fusion on the first faulty component and the first target faulty component to update the global fault database includes: fusing the first time series feature data corresponding to the first faulty component with the second time series feature data corresponding to the first target faulty component; Merging the fault location text, fault attribute text, and component name corresponding to the first faulty component into the historical fault location text, historical fault attribute text, and historical component name corresponding to the first target faulty component; The preliminary fault features and component names of each first fault component are input into a pre-trained semantic extraction model to obtain a semantic feature vector corresponding to each first fault component.

3. A wind turbine generator set early fault detection system according to claim 2, characterized in that: Performing feature fusion on the first faulty component and the first target faulty component to update the global fault database further includes: Searching for a first target fault component that matches each first fault component among a plurality of second fault components included in the global fault database, and fusing semantic feature vectors corresponding to the first target fault components into a historical semantic feature vector corresponding to the first target fault component; The second time series feature data, historical fault location text, historical fault attribute text, historical component name and historical semantic feature vector corresponding to the first target fault component are added to the global fault database to update the global fault database.

4. A wind turbine generator set early fault detection system according to claim 3, characterized in that: The fault location text, fault attribute text, and component name corresponding to the first fault component are respectively integrated into the historical fault location text, historical fault attribute text, and historical component name corresponding to the first target fault component, including: Searching for a standard component name that matches the component name of the first fault component in a preset wind turbine generator set standard fault table; Merging the standard component name of the first faulty component into the historical component name corresponding to the first target faulty component; The fault location text and the fault attribute text of the first fault component are added to the historical fault location text and the historical fault attribute text corresponding to the first target fault component.

5. A wind turbine generator set early fault detection system according to claim 4, characterized in that: Performing feature fusion on the first faulty component and the first target faulty component to update the global fault database further includes: When the first early fault detection time of the historical fault location text corresponding to the second fault component is before the preset maintenance cycle, and the number of early fault detections of the historical fault attribute text corresponding to the second fault component is less than the first preset threshold, the historical fault location text corresponding to the second fault component is deleted from the updated global fault database.

6. A wind turbine generator set early fault detection system according to claim 5, characterized in that: Matching the first faulty component with the global fault database to obtain a second faulty component that matches the first faulty component, and using the second faulty component as a first target faulty component, includes: Determine a timing correlation and a first feature similarity between each second fault component and the first fault component in the global fault database; determining a target correlation based on a timing correlation and a first feature similarity between the second faulty component and the first faulty component; A second fault component whose target correlation with the first fault component is greater than a second preset threshold among the second fault components in the global fault database is determined as a first target fault component matching the first fault component.

7. A wind turbine generator early fault detection system according to claim 6, characterized in that: Obtaining the local state of the unit corresponding to the current operating data includes: Obtaining vibration time series data and temperature time series data corresponding to the current operating data; generating a time-frequency characteristic graph corresponding to the current operation data according to the vibration time series data and the temperature time series data; The local state of the unit is generated according to the time-frequency characteristic diagram, and first time series characteristic data corresponding to the first faulty component is generated in the local state.

8. The wind turbine generator set early fault detection system according to claim 7, characterized in that: The retrieval feature includes at least one of time series feature data, text feature and keyword feature of component name; wherein the text feature includes fault location text feature and fault attribute text feature.

9. The wind turbine generator set early fault detection system according to claim 8, characterized in that: Matching the search feature with the updated global fault database to obtain a second target faulty component that matches the component to be inspected, comprising: Determine a timing matching degree between the timing characteristic data and each second fault component in the updated global fault database; Determining a semantic match between the text feature and each second fault component in the updated global fault database; Determine a second target faulty component in the updated global fault database that matches the search feature according to the timing matching degree and the semantic matching degree corresponding to the second faulty component.

10. The wind turbine generator set early fault detection system according to claim 9, characterized in that: Matching the search feature with the updated global fault database to obtain a second target faulty component that matches the component to be inspected, further comprising: When the number of keywords included in the search feature is less than or equal to a third preset threshold, determining a keyword matching degree between the keyword feature and each second fault component in the updated global fault database; When the number of keywords included in the search feature is less than or equal to a third preset threshold, a second target faulty component matching the search feature in the updated global fault database is determined according to the timing matching degree and keyword matching degree corresponding to the second faulty component.

Images