Method for constructing offshore wind turbine fault sensitivity evaluation knowledge graph

Abstract

The invention discloses a method for constructing an offshore wind turbine fault sensitivity evaluation knowledge graph. The method comprises a method for avoiding limitation caused by lack of rapid diagnosis and advanced prediction of similar faults of a wind turbine generator. The structured marine meteorological data and the fan SCADA operation data are processed, fault sensitive feature factors are extracted, and feature factors with high fault correlation are comprehensively extracted; abstraction fault sensitivity characteristic factors are used as nodes, unstructured information in operation and maintenance data is abstracted into entity nodes or attribute nodes, the relation between the nodes is established, the incidence relation of fan faults is understood and analyzed, a graphic structure is presented through a visual tool, and the fault diagnosis efficiency is improved. A user can visually carry out query and reasoning operation, and the situation that personnel lack rapid diagnosis and advance prediction on similar faults of the wind turbine generator through an SCADA system and working experience of the personnel, and limitation occurs is avoided.

Description

Technical Field

[0001] The present invention relates to a method for constructing a knowledge graph for evaluating fault sensitivity of offshore wind turbines, and relates to the technical field of offshore wind turbines. Background Art

[0002] Offshore wind turbines are installed on the sea surface and operate in extremely harsh environments such as high salt fog and high humidity, which makes the failure rate of offshore wind farms high. Therefore, it is necessary to predict the trend of wind turbine equipment failure and performance degradation in advance to take maintenance measures in time.

[0003] At present, the unit operation and maintenance personnel mainly use the SCADA system and their own work experience to troubleshoot wind turbines. However, this method has the limitation of lack of rapid diagnosis and early prediction of similar faults. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides a method for constructing a knowledge graph for evaluating fault sensitivity of offshore wind turbines.

[0005] The present invention is achieved through the following technical solutions.

[0006] The present invention provides a method for constructing a knowledge graph for evaluating fault sensitivity of offshore wind turbines, comprising:

[0007] A method to avoid the limitations of lack of rapid diagnosis and early prediction of similar faults in wind turbines.

[0008] The method comprises:

[0009] Step 1: Obtain offshore climate data, wind turbine SCADA operating status data, and wind turbine operation and maintenance data, pre-process and integrate the data to form a comprehensive offshore wind turbine equipment operation data set;

[0010] Step 2: Use machine learning algorithms to perform feature engineering on the structured offshore climate data and wind turbine SCADA operation data in the comprehensive operation dataset to extract offshore wind turbine fault-sensitive characteristic factors;

[0011] Step 3: Abstract the fault sensitivity characteristic factors into nodes. Each column of the unstructured data in the wind turbine operation and maintenance data has been stored according to the fault component, fault time, and fault cause. The content in each column cell can be directly used as the entity node or attribute node in the knowledge graph. Then analyze the relationship between the nodes and store the relationship as the edge in the knowledge graph. The nodes and edges are stored in the graph database in the form of triples.

[0012] Step 4: Use the triples stored in the graph database to build a knowledge graph, and use visualization tools to present the graph structure and provide users with query and reasoning operations.

[0013] The beneficial effects of the present invention are as follows: the present application processes structured offshore meteorological data and wind turbine SCADA operation data, extracts fault-sensitive characteristic factors, and comprehensively extracts characteristic factors with high fault correlation; uses abstract fault-sensitive characteristic factors as nodes, abstracts unstructured information in operation and maintenance data into entity nodes or attribute nodes, establishes relationships between nodes, thereby understanding and analyzing the correlation between wind turbine failures, and presents a graphical structure through a visualization tool, so that users can intuitively perform query and reasoning operations, thereby avoiding the situation in which personnel lack rapid diagnosis and advance prediction of similar wind turbine failures through the SCADA system and their own work experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a schematic diagram of the overall workflow of the present invention;

[0015] Figure 2 is a flow chart of data preprocessing and integration of the present invention;

[0016] Figure 3 is a flow chart of extracting fault sensitivity characteristic factors of offshore wind turbines according to the present invention;

[0017] Figure 4 is a flow chart of constructing a knowledge graph for evaluating fault sensitivity of offshore wind turbines according to the present invention; DETAILED DESCRIPTION

[0018] The technical solution of the present invention is further described below, but the scope of protection claimed is not limited to the description.

[0019] like Figures 1 to 4 shown.

[0020] A method for constructing a knowledge graph for evaluating fault sensitivity of offshore wind turbines in the present application includes the following steps.

[0021] Step 1: Obtain offshore climate data, wind turbine SCADA operating status data, and wind turbine operation and maintenance data, pre-process and integrate the above data to form a comprehensive offshore wind turbine equipment operation data set.

[0022] Step 1: Preprocessing and integrating data, divided into three sub-steps: Step 1.1 to Step 1.3. Figure 2 As shown, the details are as follows.

[0023] Step 1.1: Clean the raw data obtained from various data sources, including processing missing values ​​and outliers.

[0024] Step 1.2: Unify the format of the data cleaned in step 1.1 and synchronize the timestamps of different data sources to ensure that the data is consistent in the time dimension.

[0025] Step 1.3: Using the timestamp in step 1.2 as the association key, the relevant data are integrated using the data frame merging method to form a comprehensive offshore wind turbine operation data set.

[0026] Step 2: Use machine learning algorithms to perform feature engineering on the structured offshore climate data and wind turbine SCADA operation data in the comprehensive operation dataset to extract offshore wind turbine fault-sensitive characteristic factors.

[0027] Effectiveness of data integration and preprocessing: Offshore meteorological data, wind turbine SCADA operating status data, and wind turbine operation and maintenance data are unified and associated in format to form a comprehensive operation data set, which helps to improve data consistency and availability and provide a high-quality data foundation for subsequent feature engineering and fault diagnosis.

[0028] Step 2: Extracting offshore wind turbine fault sensitive characteristic factors, divided into three sub-steps: Step 2.1 to Step 2.3. Figure 3 As shown, the details are as follows.

[0029] Step 2.1: First, construct the features, perform statistical feature calculation on the signals and time series data in the comprehensive operation data set to extract their statistical features, and use the autocorrelation time domain analysis method to extract the time domain features, and use Fourier transform to perform frequency domain analysis to extract frequency domain features.

[0030] Step 2.2: Combine the statistical features in step 2.1 and the features extracted from the time domain and frequency domain with the original meteorological features and operation status features in the comprehensive operation data set to form a comprehensive feature set [X].

[0031] Step 2.3: Use the random forest (RF) algorithm to sort the importance of the fault-related features in the feature set [X], and select the feature factors with high fault correlation. The principle is to first understand the contribution of each feature on each decision tree. The difference in the Gini index before and after a feature branches on a node is the contribution value of the feature; the same method is applied to other features, and finally the change value of the Gini index of a feature is divided by the sum of the change values ​​of the Gini index of all features to obtain the normalized contribution of a feature; finally, sort the features according to the size of the contribution. The specific method is as follows:

[0032] Let k represent k characteristic factors and represent the weight of the kth characteristic factor, then we can get:

[0033]

[0034] Therefore, the change in the Gini index of feature j at node m can be calculated by the following formula:

[0035] VIM jm =GIm -GI l -GI r

[0036] Among them VIM jm Indicates the Gini index change value of feature j on node m, GI m Represents the Gini index before branching, GI l and GI r is the Gini index of the two new nodes generated after node m branches. Assuming that feature j appears in node set M in decision tree i, the change in the Gini index of feature j in the i-th decision tree is:

[0037]

[0038] Then the normalized value of the contribution of feature j is recorded as the contribution of feature j, and the calculation method is as follows:

[0039]

[0040] Where represents the contribution of feature j after normalization;

[0041] Represents the sum of the Gini index differences of all features.

[0042] The comprehensiveness of the feature engineering of this application: machine learning algorithms are used to process structured offshore meteorological data and wind turbine SCADA operation data to extract fault-sensitive characteristic factors. This method covers the extraction of statistical features, time domain features, and frequency domain features, as well as comprehensive analysis of original meteorological features and operating status characteristics, ensuring the comprehensive extraction of characteristic factors with high fault correlation.

[0043] Step 3: Abstract fault sensitivity characteristic factors into nodes. Each column of unstructured data in the wind turbine operation and maintenance data has been stored according to faulty components, fault time, and fault causes. The content in each column cell can be directly used as an entity node or attribute node in the knowledge graph. Then analyze the relationship between nodes and store the relationship as an edge in the knowledge graph. Nodes and edges are stored in the graph database in the form of triples.

[0044] Step 3: Abstract fault sensitivity characteristic factor storage, divided into three sub-steps, steps 3.1 to 3.3, such as Figure 4 As shown, the details are as follows.

[0045] Step 3.1: Take the fault characteristic factors obtained in step 2 as nodes, and combine them with the content nodes in the corresponding columns of fault components, fault time, and fault causes to form an initial node set.

[0046] Step 3.2: Perform relationship analysis on the nodes in the node set and establish edges. This operation aims to extract the relationship between nodes using the BERT model. The relationship between nodes includes "fault component", "fault time", "fault cause" and "fault feature". First, the BERT model is trained with annotated corpus containing multiple relationships (for example: "Wind turbine No. 1", "fault component", "gearbox", "fault cause", "bearing temperature is too high", "fault feature", "bearing drive end temperature is as high as 194℃", in which node 1 is "wind turbine No. 1", node 2 is "gearbox", and the relationship type between node 1 and node 2 is "fault component", indicating that there is a "fault component" relationship between "wind turbine No. 1" and "gearbox", that is, the gearbox is a faulty component of wind turbine No. 1, and the node corresponding to its attribute 1 "fault cause" is "bearing temperature is too high", and the node corresponding to attribute 2 "fault feature" is "bearing drive end temperature is as high as 194℃"). After multiple training adjustments, the trained BERT model is used to extract relationships for all nodes and establish corresponding edges.

[0047] Step 3.3: The operation of storing nodes and their corresponding relationships in the form of triples in the graph database. This operation is designed to use the import function of the Neo4j graph database and use Cypher statements to import triple data into the database. In this scenario, there are four node types, namely faulty components, fault time, fault characteristics, and fault time. The relationship types are connection relationships and belonging relationships. In Neo4j, the CREATE clause is used to create nodes and relationships. Each triple includes two nodes and the relationship between them.

[0048] The establishment and visualization of the knowledge graph of the present application: using abstract fault sensitivity characteristic factors as nodes, abstracting the unstructured information in the operation and maintenance data into entity nodes or attribute nodes, and establishing the relationship between nodes, which helps to build a more flexible and detailed knowledge graph, so as to better understand and analyze the correlation between wind turbine failures, and present the graphical structure through visualization tools, so that users can intuitively perform query and reasoning operations, improving the efficiency of fault diagnosis. The efficiency of the relationship extraction of the present application: using annotated corpus containing multiple relationships to train the model to accurately extract the relationship between nodes, this method can better handle complex relationship networks, making the constructed knowledge graph more accurate and comprehensive.

[0049] Step 4: Use the triples stored in the graph database to build a knowledge graph, and use visualization tools to present the graph structure and provide users with query and reasoning operations.

[0050] Step 4 uses triples to construct the knowledge graph, which is divided into two sub-steps: Step 4.1 to Step 4.2, as follows.

[0051] Step 4.1: After the triple data is imported, various query operations are performed using Neo4j's query language Cypher. Through query operations, the relationship path between specific feature factors is obtained, the nodes in the path are analyzed, and potential fault sensitivity feature factors are identified. This operation traverses the entire knowledge graph based on the depth-first traversal algorithm and obtains the information in it. To ensure that the system can fully traverse at each information level, the information level is divided into four information layers, namely the unit component layer, the fault mode layer, the fault feature layer, and the equipment status layer. For a clearer description, the nodes at each level are named source node R, root node G, and so on. i , child node Z i , Leaf node Y i When performing the traversal algorithm, the program will only visit each node in the path once. The traversal route is as follows: first visit the source node R to extract its written information and set the node as visited; then traverse its adjacent root node G 1 After extracting the content of the node, set it as visited; according to the identifier, access the child node Z associated with it 1 , traverse its fault characteristics to obtain information and set it as visited; then continue to visit Z 1 Neighboring node Y 1 After getting the content, set it as visited; then traverse Z in sequence 1 Y under the child node 2 To the end of Y i , get all the written content; then continue to traverse G 2 To G i , until all nodes are traversed, the fault characteristics and device status are repeatedly traversed, ensuring that all paths of child nodes are recursively traversed until the bottom-level leaf node and then the program ends to output its traversal results.

[0052] Step 4.2: Use the inference algorithm to predict the possible fault types and probabilities of the wind turbine under different environmental conditions based on the existing knowledge graph. The operation of using the inference algorithm to predict the possible fault types and probabilities of the wind turbine under different conditions. This operation is based on the fault sensitivity characteristic factors and knowledge graph information obtained in step 4.1 to establish a rule base. The rule base includes the possible fault types and corresponding probabilities under each condition. The rules may include unit components, fault modes, fault characteristics, and equipment status. The rule base, unit monitoring status conditions, and environmental conditions are used as inputs for reasoning. The reasoning process first locates the corresponding unit component nodes based on the input conditions; then, based on the information of the component nodes, the possible fault modes are inferred; then, based on the fault modes, the possible fault characteristics and equipment status are inferred; then, the probability of each fault is estimated by combining frequency information and historical data; finally, the inferred fault type and its probability distribution are output.

[0053] Flexibility of graph databases: Nodes and relationships are stored in the form of triples in graph databases, using graph database technologies such as Neo4j. This improves query efficiency and flexibility, making knowledge graphs easier to manage and update.

[0054] Our entire method comprehensively considers multiple links such as data processing, feature engineering, and knowledge graph establishment, thereby constructing a knowledge graph for offshore wind turbine fault sensitivity evaluation to improve the diagnostic accuracy and prediction ability of offshore wind turbine faults, reduce operation and maintenance management costs, and improve efficiency.

Claims

1. A method for constructing a knowledge graph for evaluating fault sensitivity of offshore wind turbines, characterized in that: include: A method to avoid the limitations of lack of rapid diagnosis and early prediction of similar faults in wind turbines.

2. The method for constructing a knowledge graph for evaluating fault sensitivity of offshore wind turbines according to claim 1, characterized in that: The method comprises: Step 1: Obtain offshore climate data, wind turbine SCADA operating status data, and wind turbine operation and maintenance data, pre-process and integrate the data to form a comprehensive offshore wind turbine equipment operation data set; Step 2: Use machine learning algorithms to perform feature engineering on the structured offshore climate data and wind turbine SCADA operation data in the comprehensive operation dataset to extract offshore wind turbine fault-sensitive characteristic factors; Step 3: Abstract the fault sensitivity characteristic factors into nodes. Each column of the unstructured data in the wind turbine operation and maintenance data has been stored according to the fault component, fault time, and fault cause. The content in each column cell can be directly used as the entity node or attribute node in the knowledge graph. Then analyze the relationship between the nodes and store the relationship as the edge in the knowledge graph. The nodes and edges are stored in the graph database in the form of triples. Step 4: Use the triples stored in the graph database to build a knowledge graph, and use visualization tools to present the graph structure and provide users with query and reasoning operations.

3. The method for constructing a knowledge graph for evaluating fault sensitivity of offshore wind turbines according to claim 2, characterized in that: The step 1 comprises: Step 1.1: Clean the raw data obtained from various data sources, including processing missing values ​​and outliers; Step 1.2: Unify the format of the data cleaned in step 1.1, and synchronize the timestamps of different data sources to ensure that the data is consistent in the time dimension; Step 1.3: Using the timestamp in step 1.2 as the association key, use the data frame merging method to integrate the relevant data to form a comprehensive offshore wind turbine operation data set.

4. The method for constructing a knowledge graph for evaluating fault sensitivity of offshore wind turbines according to claim 2, characterized in that: The step 2 comprises: Step 2.1: First, construct features, perform statistical feature calculation on the signals and time series data in the comprehensive operation data set to extract their statistical features, use the autocorrelation time domain analysis method to extract time domain features, and use Fourier transform to perform frequency domain analysis to extract frequency domain features; Step 2.2: Combine the statistical features in step 2.1 and the features extracted from the time domain and frequency domain with the original meteorological features and operation status features in the comprehensive operation data set to form a comprehensive feature set.

5. The method for constructing a knowledge graph for evaluating fault sensitivity of offshore wind turbines according to claim 4, characterized in that: The step 2 also includes step 2.3: using a random forest algorithm to sort the importance of the fault-related features in the feature set, and screening out feature factors with high fault relevance. Here’s how: Let k represent k characteristic factors and represent the weight of the kth characteristic factor, then we can get: Therefore, the change in the Gini index of feature j at node m can be calculated by the following formula: VIM jm =GI m -AND l -AND r Among them VIM jm Indicates the Gini index change value of feature j on node m, GI m Represents the Gini index before branching, GI l and GI r is the Gini index of the two new nodes generated after the node m branches; Assume that feature j appears in the node set M in decision tree i, then the change in the Gini index of feature j in the i-th decision tree is: Then the normalized value of the contribution of feature j is recorded as the contribution of feature j, and the calculation method is as follows: Where represents the contribution of feature j after normalization; Represents the sum of the Gini index differences of all features.

6. The method for constructing a knowledge graph for evaluating fault sensitivity of offshore wind turbines according to claim 2, characterized in that: The step 3 comprises: Step 3.1: Take the fault characteristic factors obtained in step 2 as nodes, and combine them with the content nodes in the corresponding columns of fault components, fault time, and fault causes to form an initial node set; Step 3.2: Perform relationship analysis on the nodes in the node set and establish edges. This operation aims to extract the relationships between nodes using the BERT model. The relationships between nodes include "faulty component", "fault time", "fault cause" and "fault characteristics". First, train the BERT model with annotated corpus containing multiple relationships. After multiple training adjustments, use the trained BERT model to extract relationships from all nodes and establish corresponding edges.

7. The method for constructing a knowledge graph for evaluating fault sensitivity of offshore wind turbines according to claim 6, characterized in that: The step 3 also includes step 3.3: an operation of storing nodes and nodes and their corresponding relationships in the form of triples in a graph database; this operation is intended to utilize the import function of the Neo4j graph database and use Cypher statements to import triple data into the database; in this scenario, there are four node types, namely, fault component, fault time, fault feature and fault time, and the relationship types are connection relationship and belonging relationship. The CREATE clause is used in Neo4j to create nodes and relationships, and each triple includes two nodes and the relationship between them.

8. The method for constructing a knowledge graph for evaluating fault sensitivity of offshore wind turbines according to claim 2, characterized in that: The step 4 comprises: Step 4.1: After the triple data is imported, use Neo4j's query language Cypher to perform various query operations. Through query operations, obtain the relationship path between specific characteristic factors, analyze the nodes in the path, and identify potential fault sensitivity characteristic factors; Step 4.2: Use the inference algorithm to predict the possible fault types and probabilities of the fan under different environmental conditions based on the existing knowledge graph.

Images