Fan fault diagnosis method and system based on dynamic fault tree full-dimension knowledge base

Abstract

The application discloses a fan fault diagnosis method based on a dynamic fault tree full-dimension knowledge base, relates to the technical field of wind power generation, and is used for solving the problem of inaccurate fan fault diagnosis of the prior art, and comprises the following steps: obtaining fan data and performing data processing, wherein the fan data comprises operation data, text data and image data; according to the fan data after data processing, entities and the relationship between the entities are extracted, a graph convolution network is used to model the entities and the relationship between the entities to obtain a fan fault tree knowledge base used for fan fault diagnosis; new fan data is obtained, and entities and the relationship between the entities are extracted; and when there are new entities and the relationship between the entities, the fan fault tree knowledge base is updated. The application further discloses a fan fault diagnosis system based on a dynamic fault tree full-dimension knowledge base. Through data integration and extraction of the relationship between entities, the application improves the fault diagnosis accuracy.

Description

Technical Field

[0001] This invention relates to the field of wind power generation technology, and more particularly to a wind turbine fault diagnosis method and system based on a dynamic fault tree full-dimensional knowledge base with automatically updated knowledge base. Background Technology

[0002] When generating electricity through wind power, the stability and reliability of wind turbine operation directly affect power generation efficiency and economic benefits. Currently, due to the complex and variable operating environment of wind turbines and their high failure rate, traditional fault diagnosis methods suffer from at least the following technical problems: over-reliance on human experience, scattered and unintegrated data, insufficient knowledge extraction capabilities, and imperfect knowledge update mechanisms.

[0003] Therefore, there is an urgent need for a method that can more accurately diagnose wind turbine faults. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, one of the objectives of this invention is to provide a wind turbine fault diagnosis method based on a dynamic fault tree full-dimensional knowledge base. This method learns from the operating data, text data, and image data in the wind turbine data to identify fault-related entities and relationships, and constructs a fault diagnosis knowledge base for wind turbine fault judgment. It also continues to learn from newly acquired wind turbine data to continuously improve the knowledge base.

[0005] One of the objectives of this invention is achieved through the following technical solution:

[0006] A wind turbine fault diagnosis method based on a dynamic fault tree full-dimensional knowledge base includes the following steps:

[0007] Acquire wind turbine data and perform data processing; the wind turbine data includes operating data, text data, and image data.

[0008] Based on the processed wind turbine data, entities and relationships between entities are extracted, and a wind turbine fault tree knowledge base for wind turbine fault diagnosis is obtained by modeling the entities and relationships between entities using a graph convolutional network.

[0009] New wind turbine data is acquired, and entities and relationships between them are extracted. When new entities and relationships exist, the wind turbine fault tree knowledge base is updated using an incremental fusion algorithm.

[0010] To ensure data accuracy and facilitate the extraction of entity and relational data, wind turbine data is acquired and processed, including:

[0011] Collect the operating data of the wind turbine, and divide the operating data into low-frequency data and high-frequency data according to the collection frequency;

[0012] Obtain text data of wind turbine maintenance records, perform entity recognition on the text data, and store the recognized data;

[0013] Acquire image data of different components or locations of the wind turbine, and extract defect features from the image data;

[0014] Based on the text data, the text data of the operation data, maintenance records, and image data are correlated, fused, and stored.

[0015] Furthermore, the process of collecting wind turbine operating data and dividing the operating data into low-frequency data and high-frequency data according to the collection frequency also includes: extracting the spectral features of high-frequency data through fast Fourier transform, storing the low-frequency data in a relational database, and storing the high-frequency data in a time-series database.

[0016] The process involves acquiring text data from wind turbine maintenance records, performing entity recognition on the text data, and storing the recognized data. This includes training a BERT model for semantic understanding based on the text data, performing Chinese word segmentation and part-of-speech tagging on the text data, and finally extracting the equipment name, fault system, and fault cause corresponding to each fault event in the text data using a BiLSTM-CRF model, thereby constructing a standardized fault database.

[0017] Acquire image data of different components or locations of the wind turbine, and extract defect features from the image data, including: performing edge detection on the image data using the Canny algorithm, and extracting defect features using a convolutional neural network.

[0018] Entities and relationships are categorized by type to enable more accurate fault analysis based on different entities / relationships. The entities include: system name, component name, operating status, special events, and event causes. The relationships between entities include attribution relationships and causal relationships.

[0019] Furthermore, a wind turbine fault tree knowledge base for wind turbine fault diagnosis is obtained by modeling the entities and the relationships between them using a graph convolutional network, including:

[0020] The processed wind turbine data is input into a graph convolutional network for optimization to obtain entity vectors.

[0021] By embedding the entity vectors output by the graph convolutional network using TransE, a mapping result between entities and their relationships is obtained, completing the knowledge representation modeling and serving as a wind turbine fault tree knowledge base. The mapping function of TransE satisfies: f(s,r,o)=||h s +h r -h o||, where s, r, and o are the subject, relation, and object, respectively, and h s h o Subject and object entity embedding vectors, h r It is a relation vector;

[0022] Obtain the expert experience rules and represent them as first-order predicate logic (FOL) to obtain the relation triples of the expert experience rules;

[0023] The relationship triples are embedded into TransE to improve the wind turbine fault tree knowledge base.

[0024] Furthermore, to simultaneously provide the probability of fault occurrence, the relation triples are embedded in TransE. After improving the wind turbine fault tree knowledge base, it also includes:

[0025] The probability of each fault occurring is calculated using Bayesian inference, and the Bayesian inference calculation satisfies the following:

[0026] Where F represents a fault event, X is an observed feature, P(F|X) is the probability of fault event F occurring given observed feature X; P(X|F) is the likelihood of observed feature X occurring when fault event F occurs; P(F) is the occurrence rate of fault event F in historical data, and P(X) is the probability of the current observed feature X occurring.

[0027] To facilitate efficient querying of the knowledge base, the wind turbine fault tree knowledge base is stored in the Neo4j graph database and queried using the Cypher query language.

[0028] When performing fault diagnosis, feature matching is performed using the wind turbine fault tree knowledge base to output the fault cause and fault probability.

[0029] Furthermore, in order to supplement the knowledge base through reasoning methods, acquire new wind turbine data, and extract entities and relationships between them, when new entities and relationships exist, the wind turbine fault tree knowledge base is updated using an incremental fusion algorithm, including:

[0030] Calculate the extracted new entities E1, E2 and the new relationship R between the entities. new The similarity between (E1, E2) and existing entities and relationships between entities, wherein the similarity S(R) is... new ,R exist The calculation of ) satisfies the formula: S(R) new ,R exist )=α·S 实体 +β·S 关系 +γ·S 属性 Where α, β, and γ are fusion weights, satisfying α + β + γ = 1, S实体 S represents the similarity between entity vectors. 关系 S represents the similarity of relations. 属性 Indicates the similarity of entity attributes;

[0031] Entity vector similarity is calculated using cosine similarity, satisfying:

[0032] in, For the vector representation of the new entity and the existing entity of entity E1;

[0033] Relation similarity is calculated through relation vector embedding and satisfies:

[0034]

[0035] Where h1 and h2 represent the vector representations of the new entities E1 and E2, r exist This represents the old relationships in the wind turbine fault tree knowledge base;

[0036] Entity attribute similarity is calculated using Jaccard similarity and satisfies:

[0037] Among them, A new A exist For entities E1 and E2, this is the set of attributes under the new and existing relations.

[0038] When the similarity is less than or equal to the threshold, the new relationship between the entities is added to the wind turbine fault tree knowledge base; otherwise, data fusion is performed.

[0039] The source of the new relationship between entities is determined based on a preset priority. If it originates from a high-priority entity, the new relationship replaces the corresponding old relationship in the wind turbine fault tree knowledge base. Otherwise, the new relationship and the corresponding old relationship in the wind turbine fault tree knowledge base are weighted and merged to satisfy the following conditions:

[0040] R confused =λR exist +(1-λ)R new , where R confused R represents the fused relational vector. exist This represents the old relation, where λ is the dynamic weighting coefficient, which is calculated using data reliability and an exponential decay factor, satisfying the following:

[0041] Among them, C exist and C new Here, δt represents the confidence levels of the old and new relations, respectively, and et represents the time difference between the corresponding data of the old and new relations. -δtIt is an exponential decay factor;

[0042] When the new wind turbine data contains qualitative knowledge, it is directly added to the wind turbine fault tree knowledge base.

[0043] To further improve the database, new wind turbine data is acquired, and entities and relationships between them are extracted. When new entities and relationships exist, the wind turbine fault tree knowledge base is updated using an incremental fusion algorithm. The process also includes: acquiring other knowledge bases and sharing knowledge with the wind turbine fault tree knowledge base, including the following steps:

[0044] Acquire other knowledge bases, which are comprehensive wind farm knowledge bases with rich data.

[0045] The difference in knowledge distribution between the two knowledge bases was calculated using the maximum mean difference method, and then normalized.

[0046] Knowledge base sharing is performed on a knowledge base adjusted according to normalization, and the knowledge base sharing satisfies the objective function:

[0047] Among them, f θ Let θ be the knowledge mapping function, and θ be the optimization parameter. Let ω be the loss function. i For the weighting coefficient, if ω i If K ≤ 1, discard the corresponding knowledge point; otherwise, migrate the knowledge point to the wind turbine fault tree knowledge base. T For wind turbine fault tree knowledge base, K S For other knowledge bases, i represents the knowledge point number, and N represents the total number of knowledge points.

[0048] The second objective of this invention is to provide a wind turbine fault diagnosis system based on a dynamic fault tree full-dimensional knowledge base, which includes a processor, a storage medium and a computer program. When the computer program is executed by the processor, it implements the above-mentioned wind turbine fault diagnosis method based on a dynamic fault tree full-dimensional knowledge base.

[0049] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0050] This invention integrates operational and textual data, transforming previously scattered data into a knowledge graph. By extracting relationships between entities, it links the data together, uncovering deep-seated connections in complex faults, thus improving fault diagnosis efficiency and accuracy. Furthermore, this invention combines image data with other data to enrich fault diagnosis information, enhancing the understanding and handling of complex faults and avoiding diagnostic errors. In addition, this invention analyzes newly acquired wind turbine data, continuously improving the wind turbine fault tree knowledge base and performing real-time optimization of the knowledge graph, resulting in highly timely fault diagnosis.

[0051] Furthermore, this invention also utilizes transfer learning to share different knowledge bases, making the wind turbine fault tree knowledge base data complete and rich, and the diagnostic results accurate. Attached Figure Description

[0052] Figure 1 This is a flowchart of the wind turbine fault diagnosis method based on a dynamic fault tree full-dimensional knowledge base, as described in Implementation Example 1.

[0053] Figure 2 This is a flowchart of the knowledge base update process in Example 2;

[0054] Figure 3 This is a schematic diagram of a wind turbine blade defect failure in Example 4;

[0055] Figure 4 This is a schematic diagram of the fault diagnosis hierarchy in Example 4;

[0056] Figure 5 This is a schematic diagram of the fault diagnosis results of a certain brand of fan in Example 4. Detailed Implementation

[0057] The present invention will now be described in more detail with reference to the accompanying drawings. It should be noted that the following description of the present invention with reference to the accompanying drawings is merely illustrative and not restrictive. Various embodiments can be combined with each other to form other embodiments not shown in the following description.

[0058] Example 1

[0059] Example 1 provides a wind turbine fault diagnosis method based on a dynamic fault tree full-dimensional knowledge base. It aims to analyze wind turbine data and extract entities and relationships, and then infer the relationship between faults and various entities and images. In this way, a knowledge-rich wind turbine fault tree knowledge base can be built without relying on traditional manual analysis.

[0060] Traditional fault diagnosis methods suffer from technical problems such as over-reliance on human experience, fragmented and unintegrated data, insufficient knowledge extraction capabilities, and imperfect knowledge update mechanisms.

[0061] Specifically, the problem with over-reliance on human experience lies in the lack of systematic knowledge support. The varying skill levels and experience of different personnel lead to significant fluctuations in the accuracy and efficiency of fault diagnosis, making it difficult to guarantee the quality and timeliness of maintenance work. Manual diagnosis takes an average of 4.2 hours and has a misjudgment rate of 18.7%.

[0062] The problem of fragmented and unintegrated data lies in the fact that equipment operation data, maintenance records, and other information are stored in a scattered manner without effective integration and correlation analysis. This makes it extremely difficult to analyze fault patterns and trends from a global perspective, and it is impossible to fully tap the value of the data to provide strong support for fault diagnosis and maintenance decisions.

[0063] The problem with insufficient knowledge extraction capabilities lies in the fact that traditional methods struggle to uncover deep-seated relationships between complex faults and cannot accurately grasp the evolution patterns of faults under different operating conditions. For example, the probability of fault occurrence, manifestations, and propagation paths vary significantly under different seasons and wind speeds, but existing technologies cannot effectively identify and analyze these differences, thus affecting the effectiveness of fault diagnosis and prevention.

[0064] The main problem with the imperfect knowledge update mechanism is that it cannot dynamically optimize the knowledge graph in real time based on new maintenance data and experience, significantly reducing the timeliness and practicality of the knowledge graph. A knowledge graph update delay exceeding 24 hours results in more than 30% of new faults not being addressed promptly. With the development of wind turbine technology and changes in the operating environment, new fault types and solutions are constantly emerging. If the knowledge graph cannot be updated in a timely manner, it will be unable to provide the latest and most effective guidance for maintenance work.

[0065] In summary, the main improvements of this embodiment lie in multimodal data fusion innovation, intelligent knowledge extraction model innovation, dynamic knowledge management innovation, and cross-wind farm knowledge sharing. Specifically, the multimodal data fusion innovation innovatively integrates conventional data with tacit knowledge, leveraging IoT and image recognition technologies to comprehensively collect wind turbine operation information, compensating for the shortcomings of traditional data collection and laying the foundation for constructing an accurate knowledge graph. A full-dimensional data collection system covering structured data (i.e., operational data), semi-structured data (i.e., maintenance records, text data), and unstructured data (i.e., image data) is established. The intelligent knowledge extraction model innovation employs a model combining attention mechanisms and graph neural networks to accurately extract entities and complex semantic relationships, improving the accuracy and depth of knowledge graph construction and truly reflecting the inherent patterns of wind turbine faults. The dynamic knowledge management innovation utilizes an incremental knowledge fusion algorithm, employing similarity matching, weighted confidence calculation, and time decay factors to achieve real-time optimization of the knowledge base. Cross-wind farm knowledge sharing utilizes transfer learning technology to promote knowledge sharing and optimization among different wind farms, improving the level of intelligence in the wind power operation and maintenance industry.

[0066] Based on the above analysis, this embodiment provides a wind turbine fault diagnosis method based on a dynamic fault tree full-dimensional knowledge base. Please refer to [link / reference]. Figure 1 As shown, it includes the following steps:

[0067] S1. Acquire wind turbine data and process the data, wherein the wind turbine data includes operating data, text data and image data;

[0068] Step S1 requires comprehensive collection of various wind turbine data, including real-time operating data, historical maintenance records, and wind turbine images. The collected operating data is then cleaned to remove noise, duplicates, and errors. For text data, natural language processing techniques are used for preprocessing such as word segmentation, part-of-speech tagging, and named entity recognition, and the data is converted into word vectors. For image data, convolutional neural networks are used to extract texture feature vectors.

[0069] Specifically, acquiring and processing wind turbine data includes:

[0070] Collect the operating data of the wind turbine, and divide the operating data into low-frequency data and high-frequency data according to the collection frequency;

[0071] The aforementioned method collects wind turbine operating data and divides it into low-frequency and high-frequency data based on the acquisition frequency. It also includes: extracting the spectral characteristics of the high-frequency data using a Fast Fourier Transform; storing the low-frequency data in a relational database; and storing the high-frequency data in a time-series database. The wind turbine operating data includes wind speed V, temperature T, amplitude A, and power output P or other required key parameters. The signal vector X representing the wind turbine's state is also included. t Represented as:

[0072] X t =[V t ,T t A t ,P t ,…], where t is the time index.

[0073] In this embodiment, the operational data is divided into low-frequency data and high-frequency data to improve data storage efficiency. Low-frequency data, such as power and temperature, is sampled at a frequency of 1Hz in this embodiment and stored in a relational database; high-frequency data, such as vibration signals, is sampled at a frequency of 10Hz or higher and stored in a time-series database.

[0074] The above-described Fast Fourier Transform calculation process includes: extracting spectral features:

[0075] Where F(f) is the frequency domain amplitude and x(t) is the time domain vibration signal. Since this is a common calculation method for Fast Fourier Transform data, its specific principles and processes will not be elaborated upon in this embodiment.

[0076] Obtain text data of wind turbine maintenance records, perform entity recognition on the text data, and store the recognized data;

[0077] The data mentioned above can be obtained periodically from the wind farm operation and maintenance management system through the data acquisition interface, and this embodiment does not limit the specific acquisition method.

[0078] The aforementioned text data processing specifically includes: acquiring text data from wind turbine maintenance records; performing entity recognition on the text data and storing the recognized data; including: training the BERT model on the text data for semantic understanding, enabling the BERT model to understand the semantics of technical terms such as "smoke detector triggered fault on the PU platform at the bottom of the tower". Chinese word segmentation and part-of-speech tagging are performed on the text data to improve the accuracy of technical term recognition; word segmentation and tagging can be implemented using the BERT-CRF (Conditional Random Field) model. Finally, each fault event E in the text data is extracted using the BiLSTM-CRF model. f Corresponding device name C f Faulty system S f Cause of the fault R f This allows for the construction of a standardized fault database, which is then used for entity recognition via a BiLSTM-CRF model. The storage method for this standardized fault database can be described as follows:

[0079] N represents the number of maintenance record entries.

[0080] Acquire image data of different components or locations of the wind turbine, and extract defect features from the image data;

[0081] The aforementioned image data can be images of key components such as wind turbine blades and towers taken periodically using high-resolution industrial cameras or drones.

[0082] Acquire image data of different components or locations of the wind turbine, and extract defect features from the image data, including: performing edge detection on the image data using the Canny algorithm, and extracting defect features using a convolutional neural network.

[0083] Edge detection is performed on the image data using the Canny algorithm, satisfying the following:

[0084]

[0085] Where I is the input image, G x G y These represent the gradients in the horizontal and vertical directions, respectively. It should be noted that traditional Canny algorithms require manual settings, which are insufficient to handle the challenges of complex wind turbine blade surface textures and varying lighting conditions. The improved Canny algorithm parameter adaptive mechanism in this embodiment can automatically calculate high and low thresholds based on local image contrast (such as the grayscale difference between the leading and trailing edges of the blade).

[0086] Defect features are extracted using convolutional neural networks to facilitate feature extraction of blade images for defects such as wear and corrosion. The main steps include:

[0087] Input adaptation: High-resolution blade images are processed by overlapping blocks to preserve micro-defect details;

[0088] Feature extraction: The dual-branch structure captures local texture (3×3 depthwise convolution) and global semantics (dilated convolution) respectively, and then fuses them through channel attention;

[0089] Multi-task output: Synchronously classify defect types and location coordinates.

[0090] Based on the text data, the text data of the operation data, maintenance records, and image data are correlated, fused, and stored.

[0091] The aforementioned associative and fused storage aims to improve the spatiotemporal relevance of data. Data matching can be achieved by combining timestamps with device IDs. For efficient retrieval of the fused data, a hash index is used for fast mapping. The core idea of ​​a hash index is to convert data of different modalities (structured, time-series, images, etc.) into hash values ​​to speed up queries. The index value H(I) is represented as:

[0092] H(I) = hash(X) t D history Through index values, data from different modalities can be efficiently retrieved and correlated for analysis. t D represents the current operating data of the wind turbine. history This contains historical fault data. The basic structure of the hash index table is shown in Table 1:

[0093] Table 1

[0094]

[0095] S2. Based on the processed wind turbine data, extract the entities and relationships between them, and use a graph convolutional network to model the entities and relationships between them to obtain a wind turbine fault tree knowledge base for wind turbine fault diagnosis.

[0096] The entities in S2 include: system name, component name, operating status, special event, and event cause; the relationships between these entities include attribution and causal relationships. System names include, for example, cooling system, transmission system, and hydraulic system; component names include, for example, bearings, gearboxes, engine compartments, and blades; operating status includes, for example, temperature, vibration, and power; special events include, for example, overheating, low oil level, and smoke alarm; and event causes include, for example, abnormality and damage.

[0097] In this embodiment, in order to improve the structural understanding and semantic modeling capabilities between entities in the knowledge base, a representation modeling method is adopted by jointly optimizing Graph Convolutional Network (GCN) and TransE.

[0098] The update method for GCN compute nodes is as follows:

[0099]

[0100] Where A is the adjacency matrix, representing the relationships between nodes in the fault tree knowledge base; D is the degree matrix, representing the number of connections for each node; H (l) The node features of the l-th layer; W (l) σ is a trainable parameter, and σ is an activation function, such as ReLU, which introduces non-linear representation capabilities.

[0101] The adjacency matrix and degree matrix mentioned above are explained below. Assume there are four nodes in the fault knowledge base: E1 (fan), E2 (cooling system), E3 (transmission system), and E4 (gearbox), where the entities have the following relationships:

[0102] (1) The fan (E1) is connected to the cooling system (E2) and the transmission system (E3);

[0103] (2) The transmission system (E2) is connected to the gearbox (E4);

[0104] (3) The transmission system (E3) is also connected to the gearbox (E4);

[0105] Then the adjacency matrix A can be expressed as:

[0106]

[0107] in:

[0108] A[i][j] = 1 indicates that there is a relationship between nodes Ei and Ej. A[i][j] = 0 indicates that there is no direct relationship between nodes Ei and Ej.

[0109] The corresponding degree matrix D is:

[0110]

[0111] The values ​​on the diagonal represent the number of connections for each node. For example, if the fan (E1) is connected to the cooling system (E2) and the transmission system (E3), then D11 = 2.

[0112] S2 Specifically, a wind turbine fault tree knowledge base for wind turbine fault diagnosis is obtained by modeling the entities and the relationships between them using a graph convolutional network, including:

[0113] The processed wind turbine data is input into a graph convolutional network for optimization to obtain entity vectors.

[0114] To model the semantic consistency between entities and relations in a low-dimensional vector space, the TransE method is introduced to perform embedding learning on the entity representations output by the GCN. By using TransE to embed the entity vectors output by the graph convolutional network, the mapping results between entities and their relations are obtained, completing the knowledge representation modeling and serving as a wind turbine fault tree knowledge base. The mapping function of TransE satisfies: f(s,r,o)=||h s +h r -h o ||, where s, r, and o are the subject, relation, and object, respectively, and h s h o Subject and object entity embedding vectors, h r The relation vectors are independent trainable parameters in the TransE model, learned through optimization in the embedding space along with the entity vectors, and used to model the semantic relationships between entities; the node features H output by GCN. (l+1) can be used as the initial embedding h of entities s ,,h o Furthermore, semantic mapping is completed through TransE to form a structure-enhanced vector representation, which improves reasoning and matching capabilities.

[0115] Obtain the expert experience rules and represent them as first-order predicate logic (FOL) to obtain the relation triples of the expert experience rules;

[0116] Because missing relationships may exist during the construction of the fault knowledge base, they need to be filled in through reasoning methods. Therefore, this embodiment introduces expert experience rules. Expert experience rules can be rules set based on experience or rules learned through model training; this embodiment does not limit this. Expert experience rules are represented as first-order predicate logic (FOL). For example, if "a blade develops cracks and the temperature rises abnormally," then "blade fatigue failure" may occur, which can be represented by first-order predicate logic (FOL) as follows:

[0117]

[0118] That is, when a wind power component x shows both cracks and an abnormal temperature rise, it can be determined that it may be at risk of fatigue failure. This type of logical rule directly establishes semantic relationships between entity attributes, which are used to add new triples (x, potential failure, fatigue).

[0119] The relation triples generated using FOL can be further embedded into TransE to expand the scope of graph reasoning. That is, embedding the relation triples into TransE improves the wind turbine fault tree knowledge base.

[0120] To further supplement and improve the knowledge base through reasoning methods, this embodiment also introduces fault probability. Specifically, after embedding the relation triples into TransE and improving the wind turbine fault tree knowledge base, it also includes:

[0121] The probability of each fault occurring is calculated using Bayesian inference, and the Bayesian inference calculation satisfies the following:

[0122] Where F represents a fault event, such as fatigue failure, X represents an observed feature, such as cracks, temperature, vibration, etc., P(F|X) is the probability of fault event F occurring given observed feature X; P(X|F) is the likelihood of observed feature X occurring when fault event F occurs; P(F) is the occurrence rate of fault event F in historical data, which can be obtained through a statistical knowledge base, and will not be elaborated in this embodiment; P(X) is the probability of the current observed feature X occurring, which can be obtained by statistically analyzing historical frequencies.

[0123] FOL inference provides the logical relationships between fault events, while Bayesian inference provides probability estimates of these events, together forming a joint deterministic and uncertain inference model. The diagnostic results provided in this embodiment can not only help relevant personnel understand potential faults, but also further understand the probability of each fault.

[0124] In addition, for the storage of the knowledge base, this embodiment uses the Neo4j graph database to store the wind turbine fault tree knowledge base and uses the Cypher query language to query the wind turbine fault tree knowledge base;

[0125] When performing fault diagnosis, feature matching is performed using the wind turbine fault tree knowledge base to output the fault cause and fault probability.

[0126] Fault diagnosis is performed from the overall system down to specific components, corresponding to the entities and relationships extracted by the method in this embodiment, ensuring accurate and precise diagnostic results down to the specific component. When a real-time fault occurs, the current fault characteristics are matched with wind turbine fault modes in the knowledge base. Based on known symptoms and combined with the fault tree knowledge base, possible causes and probabilities of the fault are provided.

[0127] The knowledge points in the wind turbine fault tree knowledge base are the basic units of the knowledge graph built in the Neo4j graph database. These include information such as the wind turbine's operating status, faulty components, and fault events. They are the core nodes and relational features in the wind turbine fault diagnosis process, supporting semantic reasoning and graph analysis; therefore, this embodiment chooses the Neo4j graph database. Fault association analysis can be performed using Neo4j Bloom, supporting interactive queries via natural language to find faulty systems, associated components, operating status, and historical events of the wind turbine. Combining semantic search capabilities with the knowledge graph structure of the wind power field, a multi-level, multi-path fault tree visualization is provided, improving engineers' efficiency in understanding the root causes of faults. Ultimately, a multi-level fault tree knowledge base and analysis and diagnostic results are formed, encompassing the wind turbine's faulty system, faulty components, operating status, fault events, and fault causes.

[0128] S3. Acquire new wind turbine data and extract the entities and relationships between them. When new entities and relationships exist, update the wind turbine fault tree knowledge base using an incremental fusion algorithm.

[0129] Example 2

[0130] Example 2 is based on Example 1, and mainly explains and describes the updating process of the wind turbine fault tree knowledge base.

[0131] New data may contain new entities or relationships, requiring a preliminary determination of whether they overlap with existing knowledge. For details, please refer to [link / reference needed]. Figure 2 As shown, new wind turbine data is acquired, and entities and relationships between them are extracted. When new entities and relationships exist, the wind turbine fault tree knowledge base is updated using an incremental fusion algorithm, including:

[0132] S31. Calculate the extracted new entities E1, E2 and the new relationship between entities R. new (E1, E2) represents the similarity between existing entities and the relationships between them, where E1 and E2 represent two entities connected by a relationship, such as "main shaft E1" and "lubrication system E2", corresponding to nodes in the knowledge graph. The similarity S(R) is... new ,R exist The calculation of ) satisfies the formula:

[0133] S(R new ,R exist )=α·S 实体 +β·S 关系 +γ·S 属性 Where α, β, and γ are fusion weights used to adjust the contribution of the three similarities to the total score, satisfying α + β + γ = 1, S 实体 S represents the similarity between entity vectors. 关系S represents the similarity of relations. 属性 α, β, and γ represent the similarity of entity attributes; they can be preset or automatically optimized and adjusted by the model, for example, by contrastive learning loss function and automatic weighting mechanism, or by fine-tuning an existing model, such as the TransE model.

[0134] Entity vector similarity is calculated using cosine similarity, satisfying:

[0135] in, The vector representation of the new entity and the existing entity corresponding to entity E1 represents its semantic features in the embedding space; the vector dimension is usually a fixed dimension, such as 100-dimensional or 300-dimensional, used to represent the contextual semantics and structural information of the entity.

[0136] It can be obtained through the TransE model, which will not be elaborated in this embodiment.

[0137] Relation similarity is calculated through relation vector embedding and satisfies:

[0138]

[0139] Where h1 and h2 represent the vector representations of the new entities E1 and E2, r exist This represents the old relationships in the wind turbine fault tree knowledge base; the above formula can measure whether the relationship is reasonable and valid.

[0140] Entity attribute similarity is calculated using Jaccard similarity and satisfies:

[0141] Among them, A new A exist This refers to the set of attributes for entities E1 and E2 under new and existing relationships, such as "component" and "system".

[0142] S321. When the similarity is less than or equal to the threshold, the new relationship between the entities is added to the wind turbine fault tree knowledge base; otherwise, data fusion is performed.

[0143] When the similarity is less than or equal to the threshold, it indicates that the data are entirely new entities and entity relationships, and therefore can be directly added to the knowledge base.

[0144] S322. Determine the source of the new relationship between entities according to the preset priority. If it comes from a high priority, replace the corresponding old relationship in the wind turbine fault tree knowledge base with the new relationship between entities.

[0145] In this embodiment, the high-priority data sources are selected as expert-annotated and high-precision sensor data.

[0146] After replacement, confidence levels can be updated to support subsequent inference and version traceability, providing controllability for model decisions. Confidence levels can be obtained from evaluation models of data sources or through human scoring mechanisms, such as annotation accuracy, historical sensor error rates, and prediction model accuracy. They are a measure of the reliability of knowledge entries and are used to guide the fusion and inference process. Confidence level updates satisfy the following conditions:

[0147]

[0148] Among them, C exist C new These represent the confidence levels for the old and new relationships, respectively; W exist W new These represent the weights of the data sources. In this embodiment, the weights are arranged from highest to lowest as follows: manual annotation > high-precision sensor > ordinary sensor > prediction model.

[0149] Otherwise, the new relationships between the entities and the corresponding old relationships in the wind turbine fault tree knowledge base are weighted and merged to satisfy:

[0150] R confused =λR exist +(1-λ)R new , where R confused This represents the fused relation representation vector, which is a weighted synthesis of the old and new relation information in the vector space. It is used to update the representation of this relation in the knowledge graph, so that subsequent embedding computation and reasoning are more accurate. (R) exist This represents the old relation, where λ is the dynamic weighting coefficient, which is calculated using data reliability and an exponential decay factor, satisfying the following:

[0151] Among them, C exist and C new Here, δt represents the confidence levels of the old and new relations, respectively, and et represents the time difference between the corresponding data of the old and new relations. -δt As an exponential decay factor, it ensures that the more recent the data is, the higher its proportion is in the fusion, ultimately realizing a dynamic weighted fusion mechanism with credibility and time validity as its core.

[0152] S323. When the new wind turbine data contains qualitative knowledge, it is directly added to the wind turbine fault tree knowledge base.

[0153] Qualitative knowledge, such as expert experience rules and descriptions of causal relationships, does not have a clear structure or numerical representation and does not directly replace existing entities or relationships. Instead, it is integrated using logical rules. New knowledge does not directly cover old knowledge but rather supplements or refines the existing knowledge base.

[0154] This embodiment employs an incremental update strategy, updating only the changed parts and using MERGE statements to insert new relationships in batches, reducing computational costs. Furthermore, indexes are created for frequently queried fields (such as fault events) to improve query efficiency.

[0155] Example 3

[0156] Example 3 is a further optimization based on Example 1 and Example 2. It further enhances the richness of the wind turbine fault tree knowledge base by improving the data-rich knowledge base, thereby improving the diagnostic accuracy.

[0157] This embodiment includes a source wind farm S and a target wind farm T, whose knowledge bases are K respectively. S and K T (i.e., the wind turbine fault tree knowledge base of Examples 1 and 2). The source wind farm S refers to a wind farm with a relatively complete knowledge base, such as a large amount of fault history, operational data, and expert annotations, whose knowledge graph is mature and of high quality. The target wind farm T refers to the wind farm for which knowledge is to be transferred. Due to differences in equipment brands, shorter operating time, or sparse data, its fault knowledge graph may be incomplete or of low quality. Because of the different data distributions, direct transfer may produce a negative transfer effect, i.e., incorrect knowledge affects the target wind farm. Therefore, distribution alignment and weight optimization strategies are needed to ensure effective knowledge transfer.

[0158] It should be noted that, in order to reduce the deviation in knowledge distribution between different brands of wind turbines, the maximum mean difference (MMD) method can be used to calculate the difference in knowledge distribution between the two wind farms, and then normalization adjustments can be made:

[0159]

[0160] Where n and m represent K respectively S and K T The number of knowledge points in the source and target wind farms, respectively; i and j represent the indices of the knowledge points in the source and target wind farms, respectively. If D MMD If it is too large, then a mapping transformation is performed on the knowledge to make K S and K T The distribution is more similar.

[0161] Specifically, the process involves acquiring new wind turbine data and extracting entities and relationships between them. When new entities and relationships exist, the wind turbine fault tree knowledge base is updated using an incremental fusion algorithm. The process also includes acquiring other knowledge bases and sharing this knowledge with the wind turbine fault tree knowledge base, comprising the following steps:

[0162] Acquire other knowledge bases, which are comprehensive wind farm knowledge bases with rich data.

[0163] The difference in knowledge distribution between the two knowledge bases was calculated using the maximum mean difference method, and then normalized.

[0164] Knowledge base sharing is performed on a knowledge base adjusted according to normalization, and the knowledge base sharing satisfies the objective function:

[0165] Among them, f θ Let θ be the knowledge mapping function, and θ be the optimization parameter. The loss function, such as the MSE error, ω i For the weighting coefficient, if ω i If K ≤ 1, discard the corresponding knowledge point; otherwise, migrate the knowledge point to the wind turbine fault tree knowledge base. T For wind turbine fault tree knowledge base, K S For other knowledge bases, i represents the knowledge point number, and N represents the total number of knowledge points.

[0166] Weighting coefficient ω i The calculation satisfies:

[0167] P(x i |k T ), P(x i |K S ) represents knowledge point x i The probability density of occurrence in the source wind farm and the target wind farm respectively reflects their degree of fit in their respective knowledge distributions.

[0168] In summary, the methods in Examples 1 to 3 innovatively integrate conventional data with tacit knowledge, leveraging IoT and image recognition technologies to comprehensively collect wind turbine operation information, overcoming the shortcomings of traditional data collection and laying the foundation for constructing an accurate knowledge graph. A multi-dimensional data collection system covering structured data (operating parameters), semi-structured data (maintenance records), and unstructured data (images) is established. A model combining attention mechanisms and graph neural networks is employed to accurately extract entities and complex semantic relationships, improving the accuracy and depth of knowledge graph construction and truly reflecting the inherent patterns of wind turbine faults. An incremental knowledge fusion algorithm, using similarity matching, weighted confidence calculation, and time decay factors, achieves real-time optimization of the knowledge base. Transfer learning technology is utilized to promote knowledge sharing and optimization among different wind farms, enhancing the level of intelligence in the wind power operation and maintenance industry.

[0169] Example 4

[0170] Example 4 is a specific experiment corresponding to the wind turbine fault diagnosis method based on the dynamic fault tree full-dimensional knowledge base of Examples 1 to 4.

[0171] This embodiment of the experiment selected 75 wind turbine units (numbered #01 to #75) of 4 brands (A, B, C, D) from a coastal wind farm for the experiment.

[0172] First, comprehensive data collection is performed, and the system is integrated with the SCADA system to collect real-time operational data. The specific data is as follows:

[0173] The generator bearing temperature is measured with an accuracy of 0.1℃.

[0174] Gearbox oil temperature, with a measurement accuracy of 0.1℃;

[0175] The pitch motor temperature is measured with an accuracy of 0.1℃.

[0176] The X-axis and Y-axis vibration amplitudes of the nacelle were measured with an accuracy of 0.1m.

[0177] Instantaneous wind speed, with a measurement accuracy of 0.1 m / s;

[0178] Instantaneous power, with a measurement accuracy of 0.1kW.

[0179] Then, maintenance records were collected, and historical fault data were extracted from the wind farm's troubleshooting and maintenance records. There were 133 fault maintenance records in the past year. The BERT-CRF model was used for entity recognition to extract the faulty system, faulty component, faulty event, and fault cause.

[0180] Table 2 shows a sample of the troubleshooting and maintenance record.

[0181] Table 2

[0182]

[0183] The extracted results are:

[0184] Faulty system: Cooling system;

[0185] Faulty component: Generator;

[0186] Fault event: Overheating of the generator drive end bearing;

[0187] Possible causes of failure: contactor malfunction, wiring malfunction, motor damage.

[0188] This experiment also included image data acquisition, using an industrial drone equipped with an RGB+infrared dual-mode camera to inspect wind turbine blades. Gaussian filtering, histogram equalization, and the Canny algorithm were used to preprocess the wind turbine blade images. Subsequently, a deep learning convolutional neural network (CNN) was used to detect defects such as wear and corrosion in the blade images. The results are as follows: Figure 3 As shown, the small squares (indicated by the arrows) show the detected blade defects.

[0189] By organizing the collected data across all dimensions, a wind turbine fault knowledge graph was constructed using the Neo4j graph database. This resulted in a final fault tree knowledge base that includes faulty systems, faulty components, operating states, fault events, and fault causes for wind turbine units. (See attached document.) Figure 4 Category hierarchy diagram.

[0190] By combining knowledge graphs and causal reasoning, the system automatically analyzes possible causes of malfunctions.

[0191] For example, if a "pitch communication interruption timeout fault" is detected in real time, the diagnostic result is as follows: Figure 5 As shown.

[0192] Next, the newly collected data is analyzed using an incremental update algorithm. For example, if a new fault, "pitch communication interruption timeout fault," is encountered, maintenance personnel will investigate based on the system's diagnostic results and find that the cause is still "contactor fault." The similarity S is calculated to be 0.9, indicating a high matching degree and unchanged fault tree relationship; therefore, the relevant probabilities only need to be updated.

[0193] The fault distribution difference between wind turbine brands A and B was calculated using the MMD method, with a result of 0.15, indicating suitability for direct transfer. This resulted in a comprehensive fault tree knowledge base for different wind turbine brands at the wind farm.

[0194] Example 5

[0195] Embodiment 5 of the present invention also provides a wind turbine fault diagnosis system based on a dynamic fault tree full-dimensional knowledge base, including a processor, a storage medium, and a computer program. When the computer program is executed by the processor, it implements the wind turbine fault diagnosis method based on a dynamic fault tree full-dimensional knowledge base described in Embodiments 1 to 4. The method includes:

[0196] Acquire wind turbine data and perform data processing; the wind turbine data includes operating data, text data, and image data.

[0197] Based on the processed wind turbine data, entities and relationships between entities are extracted, and a wind turbine fault tree knowledge base for wind turbine fault diagnosis is obtained by modeling the entities and relationships between entities using a graph convolutional network.

[0198] New wind turbine data is acquired, and entities and relationships between them are extracted. When new entities and relationships exist, the wind turbine fault tree knowledge base is updated using an incremental fusion algorithm.

[0199] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause an electronic device (which may be a mobile phone, personal computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0200] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

Claims

1. A wind turbine fault diagnosis method based on a dynamic fault tree full-dimensional knowledge base, characterized in that, Includes the following steps: Acquire wind turbine data and perform data processing; the wind turbine data includes operating data, text data, and image data. Based on the processed wind turbine data, entities and relationships between entities are extracted, and a wind turbine fault tree knowledge base for wind turbine fault diagnosis is obtained by modeling the entities and relationships between entities using a graph convolutional network. Acquire new wind turbine data and extract the entities and relationships between them. When new entities and relationships exist, update the wind turbine fault tree knowledge base using an incremental fusion algorithm, including: Calculate the extracted new entity , New relationships between entities ( , The similarity between the entity and existing entities and the relationships between entities, the similarity The calculation satisfies the formula: ,in, , , To integrate weights, satisfy , Indicates the similarity between entity vectors. Indicates the similarity of relationships. Indicates the similarity of entity attributes; Entity vector similarity is calculated using cosine similarity, satisfying: ,in, For entities The vector representations of the new entity and the existing entity; Relation similarity is calculated through relation vector embedding and satisfies: ， ， ,in, and Represents a new entity , The vector representation of , This represents the old relationships in the wind turbine fault tree knowledge base; Entity attribute similarity is calculated using Jaccard similarity and satisfies: ,in, , For entities , For the set of attributes under newly added relations and existing relations; When the similarity is less than or equal to the threshold, the new relationship between the entities is added to the wind turbine fault tree knowledge base; otherwise, data fusion is performed. The source of the new relationship between entities is determined based on a preset priority. If it originates from a high-priority entity, the new relationship replaces the corresponding old relationship in the wind turbine fault tree knowledge base. After replacement, the confidence level is updated, and the update satisfies the following conditions: ， in, These represent the confidence levels for the old and new relationships, respectively. These represent the weights of the data sources, respectively. Otherwise, the new relationships between the entities and the corresponding old relationships in the wind turbine fault tree knowledge base are weighted and merged to satisfy: ,in, The fused relation representation vector represents the relationship. Indicates a previous relationship. The dynamic weighting coefficients are calculated based on data reliability and exponential decay factors, and satisfy the following: ,in, and These represent the confidence levels for the old and new relationships, respectively. The time difference between the old and new data. It is an exponential decay factor; When the new wind turbine data contains qualitative knowledge, it is directly added to the wind turbine fault tree knowledge base.

2. The wind turbine fault diagnosis method based on a dynamic fault tree full-dimensional knowledge base as described in claim 1, characterized in that, Acquire wind turbine data and perform data processing, including: Collect the operating data of the wind turbine, and divide the operating data into low-frequency data and high-frequency data according to the collection frequency; The text data of the wind turbine maintenance record is obtained, entity recognition is performed on the text data, and the recognized data is stored. Acquire image data of different components or locations of the wind turbine, and extract defect features from the image data; Based on the text data, the text data of the operation data, maintenance records, and image data are correlated, fused, and stored.

3. The wind turbine fault diagnosis method based on a dynamic fault tree full-dimensional knowledge base as described in claim 2, characterized in that, Collect wind turbine operating data and divide the operating data into low-frequency data and high-frequency data according to the collection frequency. It also includes: extracting the spectral features of high-frequency data through fast Fourier transform, storing the low-frequency data in a relational database, and storing the high-frequency data in a time-series database. The process involves acquiring text data from wind turbine maintenance records, performing entity recognition on the text data, and storing the recognized data. This includes training a BERT model for semantic understanding based on the text data, performing Chinese word segmentation and part-of-speech tagging on the text data, and finally extracting the equipment name, fault system, and fault cause corresponding to each fault event in the text data using a BiLSTM-CRF model, thereby constructing a standardized fault database. Acquire image data of different components or locations of the wind turbine, and extract defect features from the image data, including: performing edge detection on the image data using the Canny algorithm, and extracting defect features using a convolutional neural network.

4. The wind turbine fault diagnosis method based on a dynamic fault tree full-dimensional knowledge base as described in claim 1, characterized in that, The entities include: system name, component name, operating status, special events, and event causes; the relationships between the entities include attribution relationships and causal relationships.

5. The wind turbine fault diagnosis method based on a dynamic fault tree full-dimensional knowledge base as described in claim 1 or 4, characterized in that, A wind turbine fault tree knowledge base for wind turbine fault diagnosis is obtained by modeling the entities and the relationships between them using graph convolutional networks, including: The processed wind turbine data is input into a graph convolutional network for optimization to obtain entity vectors. By embedding the entity vectors output by the graph convolutional network using TransE, a mapping result between entities and their relationships is obtained, thus completing the knowledge representation modeling and serving as a wind turbine fault tree knowledge base. The mapping function of TransE satisfies the following: ,in, , , They are subject, relation, and object, respectively. , Entity embedding vectors for subject and object, It is a relation vector; Obtain the expert experience rules and represent them as first-order predicate logic (FOL) to obtain the relation triples of the expert experience rules; The relationship triples are embedded into TransE to improve the wind turbine fault tree knowledge base.

6. The wind turbine fault diagnosis method based on a dynamic fault tree full-dimensional knowledge base as described in claim 5, characterized in that, After embedding the relation triples into TransE to improve the wind turbine fault tree knowledge base, it also includes: The probability of each fault occurring is calculated using Bayesian inference, and the Bayesian inference calculation satisfies the following: ,in, Indicates a fault event. For observational features, To observe given features Below, fault events The probability of occurrence; In the event of failure Observational characteristics during occurrence The possibility; For fault events in historical data The incidence rate, Current observation features The probability of its occurrence.

7. The wind turbine fault diagnosis method based on a dynamic fault tree full-dimensional knowledge base as described in claim 1, characterized in that, The wind turbine fault tree knowledge base is stored using the Neo4j graph database, and the wind turbine fault tree knowledge base is queried using the Cypher query language. When performing fault diagnosis, feature matching is performed using the wind turbine fault tree knowledge base to output the fault cause and fault probability.

8. The wind turbine fault diagnosis method based on a dynamic fault tree full-dimensional knowledge base as described in claim 1, characterized in that, Acquire new wind turbine data and extract entities and relationships between them. When new entities and relationships exist, update the wind turbine fault tree knowledge base using an incremental fusion algorithm. The process also includes: acquiring other knowledge bases and sharing knowledge with the wind turbine fault tree knowledge base, including the following steps: Acquire other knowledge bases, which are comprehensive wind farm knowledge bases with rich data. The difference in knowledge distribution between the two knowledge bases was calculated using the maximum mean difference method, and then normalized. Knowledge base sharing is performed on a knowledge base adjusted according to normalization, and the knowledge base sharing satisfies the objective function: ,in, For knowledge mapping function, To optimize parameters, For loss function, For the weighting coefficient, if If the value is ≤1, discard the corresponding knowledge point; otherwise, migrate the knowledge point to the wind turbine fault tree knowledge base. For wind turbine fault tree knowledge base, For other knowledge bases, Indicates the knowledge point number. This indicates the total number of knowledge points.

9. A wind turbine fault diagnosis system based on a dynamic fault tree full-dimensional knowledge base, comprising a processor, a storage medium, and a computer program, characterized in that, When the computer program is executed by the processor, it implements the wind turbine fault diagnosis method based on a dynamic fault tree full-dimensional knowledge base as described in any one of claims 1 to 8.

Images