A knowledge graph-based wind turbine abnormal knowledge association method and device and medium

Abstract

The application discloses a kind of wind turbine abnormal knowledge association method, device and medium based on knowledge graph, comprising: based on wind power industry operation and maintenance demand, carry out distributed multi-source heterogeneous data and knowledge collection;OPC UA information model of wind power equipment node is established;Through prompt word engineering, industrial multi-modal industrial big model is guided to node data to be normalized and be identified, complete entity extraction and attribute extraction, carry out relationship extraction and representation, to import graph database and construct abnormal knowledge graph;Optimize and update the semantic connection between node entity, complete knowledge merging and processing, dynamically update abnormal knowledge graph;Face abnormal knowledge application scene, based on abnormal knowledge graph mining entity association carries out knowledge reasoning.The application realizes the multi-modal knowledge graph construction method based on graph structure learning and fine-tuning multi-modal industrial big model, explores the correlation of different abnormalities, constructs the knowledge base rich in abnormal semantics, and ensures correct handling of industrial abnormal faults.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of artificial intelligence, and particularly relates to a wind turbine abnormal knowledge association method and device based on a knowledge graph and a medium. BACKGROUND

[0002] As a core device of renewable energy, wind turbines are widely used in remote areas (such as offshore wind farms or high-altitude mountainous areas), and their operation stability directly affects the reliability and economic benefits of the power grid. A single wind turbine is worth hundreds of thousands of yuan, and the downtime loss is huge. A 5MW unit can lose more than 100,000 yuan per day. During the operation of the wind turbine, multiple source heterogeneous data are involved, including SCADA system real-time parameters (such as wind speed, power output), vibration sensor data, thermal imaging monitoring, lubricating oil metal particle counting, and maintenance logs. These data are complex in dimension, large in size, and highly correlated in fault mode, such as gearbox abnormalities that can cause generator cascading failures. However, the wind farm environment is harsh, and the equipment is scattered, making data acquisition and integration face problems such as communication delay and data missing, making abnormal diagnosis and predictive maintenance a pain point in the industry. In the practice of fault management of wind turbines, existing methods mainly rely on manual experience or isolated systems (such as SCADA threshold alarms), resulting in scattered and lack of systematic abnormal knowledge, and the problems of data fragmentation, low knowledge utilization rate, and insufficient dynamic response. Therefore, more effective methods are needed to manage and apply fault data. To overcome the above limitations, the introduction of knowledge graph technology is a key solution. Knowledge graph can construct a semantic network of multiple source heterogeneous data (such as entityizing vibration data, maintenance records, and supply chain information into nodes), and through graph structure learning (such as graph convolution network), it can mine the implicit association between faults (such as the causal chain between main shaft imbalance and generator current abnormality). And with the help of industrial multi-modal large models, knowledge graphs can be constructed efficiently, fully leveraging the advantages of knowledge graphs in building knowledge networks and displaying knowledge associations, thereby addressing the unique characteristics of wind turbine scenarios, fieldizing general knowledge graph methods, and solving the core problems of data silos and knowledge fragmentation, providing an intelligent operation and maintenance framework for the wind power industry. SUMMARY

[0003] The present application aims to overcome the deficiencies of the prior art and provide a wind turbine abnormal knowledge association method, device and medium based on a knowledge graph.

[0004] The purpose of the present application is achieved by the following technical solutions:

[0005] A wind turbine abnormal knowledge association method based on a knowledge graph, comprising:

[0006] Step 1: Based on the operation and maintenance requirements of the wind power industry, distributed multi-source heterogeneous data and knowledge are collected;

[0007] Step 2, establish the OPC UA information model of the wind power equipment node, determine the general attributes and reference types of the node;

[0008] Step 3, guide the industrial multi-modal industrial large model to perform standardized description and identification of node data, complete entity extraction and attribute extraction, and utilize a graph convolution network to perform relationship extraction and representation, so as to import a graph database to construct an abnormal knowledge graph;

[0009] Step 4, use a graph convolution network and a long short-term memory network to optimize and update the semantic connection between node entities, complete knowledge merging and processing, and dynamically update the abnormal knowledge graph;

[0010] Step 5, use scenarios facing abnormal knowledge, and perform knowledge reasoning based on entity association mining of the abnormal knowledge graph.

[0011] Further, the step 1 comprises:

[0012] Step 1.1, multi-source knowledge data integration: the data sources include internal data sources and external data sources of wind power equipment operation and maintenance enterprises, the internal data sources include device real-time operation data, system operation and maintenance logs, and internal document materials, and the external data sources include industry standards and public data sets, academic literature and patents;

[0013] Step 1.2, data knowledge preprocessing: according to data quality, construct a feature engineering to perform average value replacement, data fitting and data rejection processing on abnormal data and missing data, and reasonably inspect knowledge sources to remove low-quality information sources.

[0014] Further, the step 2 comprises:

[0015] Step 2.1, define the OPC UA information model: the node types of the OPC UA information model include objects, object types, variables, variable types, data types, reference types, methods, and views, the general attributes of the OPC UA information model include node identification, node type, browse name, display name, description, write mask, and user write mask, and the additional attributes of the OPC UA information model include whether it is abstract, whether it is symmetric, and reverse name.

[0016] Step 2.2, construct the OPC UA information model: the establishment of the OPC UA information model includes four steps, namely, demand acquisition, definition type model, instantiation information model, and export XML and CSV documents.

[0017] Further, the step 3 comprises:

[0018] Step 3.1, standardize the OPC UA information model of the wind power equipment: use the multi-modal industrial large model with industrial data fine-tuning to preprocess the OPC UA information model in step 2, clean up unreasonable structured information, and enhance the embedded reasoning ability in the form of thought chain by combining the collected knowledge data to expand the relationship between different individuals;

[0019] Step 3.2, entity extraction and attribute extraction: based on various prompt word engineering design methods, including zero-shot learning, small sample learning, role prompt, and thought chain, guide the multi-modal industrial large model to standardize the OPC UA information model into an ontology constructed under the attribute graph model, complete entity extraction and attribute extraction;

[0020] Step 3.3, form knowledge representation: based on the relationship graph convolutional neural network, extract and represent the relationship on the preprocessed OPC UA information model data.

[0021] Further, the step 4 includes:

[0022] Step 4.1, knowledge merging and processing: based on the graph convolutional neural network, complete the knowledge merging and processing process, including link prediction, knowledge reasoning, and entity disambiguation, to form a standard knowledge representation;

[0023] Step 4.2, dynamic evolution of knowledge graph construction: use the graph convolutional neural network to capture the dependency relationship of the graph, use the long short-term memory network to learn and evolve the parameter matrix of different layers of the graph convolutional network at each time, and use knowledge embedding to construct the dynamic evolution of the knowledge graph.

[0024] Further, the link prediction in step 4.1 is based on the framework of the original knowledge graph-graph neural network encoder-decoder-knowledge graph link prediction; the graph convolutional neural network uses the graph neural network to implement the link prediction task, explicitly represents the process of updating the information of a single node in the knowledge graph, collects information from adjacent nodes, aggregates and updates the representation to the node, and iteratively updates through downstream tasks until the node vector reaches a fixed point. After learning the abnormal knowledge graph, use DisMult decoder to generate a possibility score for each potential edge in the graph, and finally select the most possible edge prediction;

[0025] The knowledge reasoning in step 4.1 includes: the graph convolutional neural network uses the learned entity embedding to predict new triples; by calculating the scores of all possible tail entities, and selecting the entity with the highest score as the prediction result;

[0026] The graph convolutional neural network in step 4.1 achieves entity disambiguation through the following process:

[0027] Each entity mention in the text is linked to an entity in the knowledge graph by calculating the similarity between the context vector of the entity mention and the embedding vector of the entity in the knowledge graph, and according to the similarity score, the most likely entity is selected as the disambiguation result of the entity mention.

[0028] Further, the step 5 comprises:

[0029] The knowledge graph reasoning application of abnormal knowledge association of a wind turbine includes intelligent retrieval of fault knowledge, identification of repetitive faults, role-based fault data analysis, effectiveness analysis of corrective measures, and predictive fault maintenance.

[0030] The application also provides a wind turbine abnormal knowledge association device based on a knowledge graph, comprising one or more processors for implementing a wind turbine abnormal knowledge association method based on a knowledge graph as described above.

[0031] The application also provides a readable storage medium having a program stored thereon, which, when executed by a processor, implements a wind turbine abnormal knowledge association method based on a knowledge graph as described above.

[0032] Compared with the prior art, the application has the following beneficial effects:

[0033] The application implements a multi-modal knowledge graph construction method based on graph structure learning and a multi-modal industrial large model, explores the correlation of different abnormalities in the wind turbine operation and maintenance scenario, constructs a knowledge base rich in abnormal semantics, and proposes an abnormal feature expression technology based on knowledge graph reinforcement, increases the probability of correctly handling faults under abnormal conditions, prolongs the service life of wind power equipment, and solves the core problems of data islanding and knowledge fragmentation, providing a landable intelligent operation and maintenance framework for the wind power industry. BRIEF DESCRIPTION OF DRAWINGS

[0034] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0035] Figure 1 is an OPC UA information model modeling flowchart in a wind turbine abnormal knowledge association method based on a knowledge graph of the application.

[0036] Figure 2 is an abnormal knowledge graph construction flowchart in a wind turbine abnormal knowledge association method based on a knowledge graph of the application.

[0037] Figure 3 FIG. 1 is a structural schematic diagram of an abnormal knowledge association device based on a knowledge graph for a wind turbine according to an embodiment of the present application. DETAILED DESCRIPTION

[0038] In order to better understand the technical solutions of the present application, the embodiments of the present application are described in detail below with reference to the drawings.

[0039] It should be clear that the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor belong to the scope of protection of the present application.

[0040] The terms used in the embodiments of the present application are only for the purpose of describing the specific embodiments, and are not intended to limit the present application. The singular forms "a", "said" and "the" used in the embodiments of the present application and the appended claims are also intended to include the plural forms, unless the context clearly indicates otherwise.

[0041] The abnormal knowledge association method based on the knowledge graph for the wind turbine provided by the embodiments of the present application, as shown in FIG. 1, comprises the following steps: Figure 2

[0042] Step 1: Based on the operation and maintenance business requirements of the wind power industry, distributed multi-source heterogeneous data acquisition is performed. Specifically, it includes:

[0043] ​Step 1.1, Multi-source knowledge data integration: Data support can be mainly divided into two aspects: internal data sources of wind power equipment operation and maintenance enterprises and external data sources. The internal data sources of wind power equipment operation and maintenance enterprises include equipment real-time operation data, system operation and maintenance logs, and internal document materials. The equipment real-time operation data mainly come from sensors, and the original SCADA data set collects more than 75 different variables. Based on the excitation knowledge of direct-drive wind power groups, 25 effective variables are selected, including hub speed, hub angle, blade 1 angle, blade 2 angle, blade 3 angle, variable pitch motor 1 current, variable pitch motor 2 current, variable pitch motor 3 current, x-direction vibration value, y-direction vibration value, nacelle weather station wind speed, wind direction absolute value, variable frequency converter grid side active power, wind tower environment temperature, nacelle temperature, blade 1 motor box temperature, blade 2 motor box temperature, blade 3 motor box temperature, blade 1 variable pitch motor temperature, blade 2 variable pitch motor temperature, blade 3 variable pitch motor temperature, blade 1 variable frequency converter box temperature, blade 2 variable frequency converter box temperature, blade 3 variable frequency converter box temperature, and wind tower environment temperature. The system operation and maintenance logs include fault types (electrical fault, control system fault, variable pitch fault, yaw fault, hydraulic fault, drive system fault, tower fault, generator fault, and nacelle fault), fault occurrence time, equipment operation data records three days before the fault occurs, fault solving method (replacement of parts, parameter calibration), and technical personnel identity (maintenance team information), forming a fault attribution chain. The internal document materials include technical documents (instrument and equipment instruction manual, design drawings, providing equipment function specification and safety threshold) and equipment maintenance guide (recording fault solving path and other core operation and maintenance knowledge, supporting abnormal working condition compliance judgment). The external data sources mainly include industry standards and public data sets, academic literature and patents, and supply chain data. Industry standards and public data sets such as fault accident database as a core data source, integrate fault reporting description (such as abnormal vibration, temperature sudden rise), processing method (shutdown for maintenance, parameter adjustment), and zero report (root cause analysis document), extract abnormal association knowledge (fault-phenomenon-solution triplets) from them, in addition to open databases such as safety production accident cases published by relevant institutions, and equipment reliability statistical reports shared by industry associations, which can supplement macro risk patterns. Academic literature and patents mainly extract advanced algorithms (such as deep learning-based anomaly detection models) and fault diagnosis logic (such as Bayesian network causal reasoning) from papers and patents to enhance the intelligent ability of the knowledge graph. The supply chain data associates supplier qualifications (such as component quality inspection reports) and raw material quality data (such as metal material hardness test values) to analyze the influence path of supply chain problems (such as poor quality bearings) on equipment abnormalities.

[0044] Step 1.2, data knowledge preprocessing: Since many data, especially text information, are filled by artificial, there are a large number of non-standard filling and non-uniform negative signs, it is necessary to construct feature engineering to replace the average value of abnormal data, missing data, data fitting and data rejection and other conventional processing, to ensure that the data meet the requirements, to reasonably test the knowledge source, to remove low-quality information sources, the data preprocessing rules include: filtering out the points with power less than or equal to zero, wind turbine limited operation and affected by switching transition; filter out the power outliers. Feature engineering is a known technology for data processing in machine learning, and will not be described here.

[0045] Step 2, establish the OPC UA information model of the wind power equipment node, determine the general attributes and reference types of the node, and model the node instance facing the equipment, as shown in Figure 1 Specifically, it includes:

[0046] Step 2.1, define the OPC UA information model: traditional OPC can only provide data information, and cannot reflect semantic association. While the OPC UA specification can not only provide data information, but also provide effective data semantics, such as the OPC UA information model not only displays the attribute information of the temperature sensor device, but also allows the type hierarchy supported by the device to be exposed. OPC UA information modeling uses object-oriented technology, including type hierarchy and inheritance; type information is exposed to the outside, and can be accessed in the same way as accessing instances; a full mesh node network is used, which allows information to be connected in various ways; type hierarchy and reference type between nodes are extensible; OPC UA has no restrictions on how to model information, in order to establish appropriate models for the data provided; OPC UA information modeling is always done on the server side.

[0047] Specifically, the node types of the OPC UA information model in the wind turbine operation and maintenance field include Object, ObjectType, Variable, VariableType, DataType, ReferenceType, Method, and View. The seven common attributes of OPC UA nodes include NodeId, NodeClass, BrowseName, DisplayName, Description, WriteMask, and UserWriteMask. The additional attributes of the OPC UA information model include IsAbstract, Symmetric, and InverseName. The main entity information includes Fault, Product, Manufacture, and Solution, and the main attribute information includes ProductName, Product No, Product Category, Fault Desc, Reporter, Fault Category, Fault Cause, Craft No, Craft Version, Process Type, Process Detail, and Processor.

[0048] Step 2.2, building the OPC UA information model: modeling mainly includes four steps: requirement acquisition, defining type model, instantiating information model, and exporting XML and CSV documents.

[0049] Taking the motor device as an example, first, the relevant information of the motor needs to be obtained, including: motor basic information, fault information, exclusion method and processing method, and associated product information. The motor basic information includes motor model, manufacturer, serial number, etc. The fault information includes fault type, fault description, fault occurrence time, and sensor parameter log before and after the fault. The exclusion method and processing method include solutions and steps for different faults. The associated product information includes other device information used with the motor, such as sensors, controllers, etc.

[0050] Define the object types (ObjectType) and variable types (VariableType) related to the motor and its fault monitoring.

[0051] Object Types (ObjectType): MotorType (basic properties and methods of the motor); Parameters (motor configuration parameters); Methods (motor operations such as start and stop); Faults (motor fault information); Troubleshooting (troubleshooting methods and handling); AssociatedProducts (associated product information).

[0052] Variable Types (VariableType): StatusVariableType (motor status variable type); ConfigurationVariableType (motor configuration parameter variable type); FaultInfoVariableType (fault information variable type); TroubleshootingVariableType (troubleshooting methods and handling variable type).

[0053] Next, instantiate the information model, instantiate specific motor objects, and include fault monitoring-related properties and methods.

[0054] MotorInstance (specific motor instance), inherits from MotorType (motor type). The instance contains Status (status): instantiates StatusVariableType, indicating the current status of the motor; Configuration (configuration): instantiates ConfigurationVariableType, containing specific configuration parameters of the motor; Faults (faults): FaultType (fault type), FaultDescription (fault description), and FaultTime (fault occurrence time); Troubleshooting (troubleshooting): Resolution (solution) and Steps (solution steps). In addition, there is AssociatedProducts (associated product information).

[0055] Finally, export XML and CSV documents using tools, use OPC UA modeling tools (such as UAmodeler or opcua-modeler) to design and export XML and CSV documents, which will serve as data sources for implementing instantiated information.

[0056] It is also necessary to bind the data source, bind the exported XML and CSV documents with the actual motor data source, and ensure that the OPC UA server can correctly read and write the state, configuration parameters, and fault information of the motor.

[0057] Finally, testing and verification are performed to test the interaction between the OPC UA client and the motor, ensuring that all properties and methods work correctly and that the data interaction process meets expectations, especially the fault monitoring and handling part.

[0058] In practical applications, the device object adjustment type model can be adjusted according to the actual application. For a variable pitch motor 1, the basic attributes can be obtained: variable pitch motor model (such as PITCH-5MW-V3), offshore corrosion level (IEC 61400-22), gear box reduction ratio (1:98.7); fault monitoring parameters: blade angle command deviation (> 2° alarm), winding temperature gradient (> 5℃ / min), backup power voltage drop (<18VDC); environmental related data: real-time wind speed (0-25m / s), atmospheric humidity (>90% icing risk), salt spray concentration (coastal wind field specific); grid interaction indicators: low voltage ride through state (LVRT Active), active power change rate (dP / dt<-0.5MW / s).

[0059] Step 3, through the prompt word engineering, guide the industrial multi-modal industrial large model to standardize the description and identification of node data, complete entity extraction and attribute extraction, use graph convolution network to extract and represent the relationship, and import the graph database to construct the abnormal knowledge graph. Specifically, it includes:

[0060] Step 3.1, standardize the information model: the main steps include data cleaning and standardization processing of OPC UA modeling information, and exploring the hidden association between OPC UA modeling objects, so as to lay a foundation for subsequent entity extraction and attribute extraction, relationship extraction and identification, and import into the graph database. The specific implementation steps will be described next.

[0061] Specifically, the preprocessing of OPC UA modeling information is mainly based on the powerful natural language and multi-modal processing capability of the multi-modal industrial large model. First, define a set of specifications and standards, which should include:

[0062] Data naming rules: unified naming conventions, such as using camel case or underscore naming.

[0063] Data type definition: clearly define the specific requirements of each data type, such as integer, floating point, string, etc.

[0064] Units and dimensions: use unified units and dimensions, such as temperature using Celsius or Fahrenheit.

[0065] Encoding rules: Define clear encoding rules for data of encoding type, such as enumeration values.

[0066] Data format: Define a unified format for data such as dates and times.

[0067] Next, use the data cleaning function of the multi-modal industrial large model to clean the non-standard data. Multi-modal industrial large models can be self-developed or from the market, such as Qizhi Kongming Industrial Large Model, Cosmo-GPT, and Industry-GPT. Data cleaning mainly includes removing invalid data, formatting, and data type conversion. Remove invalid data: Identify and remove invalid data such as null values and error values. Format uniformity: Convert data of different formats to a unified format. Data type conversion: Convert data to the correct data type.

[0068] Then use the multi-modal industrial large model to verify the data to ensure its accuracy and consistency. Mainly including range check, uniqueness check, integrity check. Range check: Check if the data is within a reasonable range. Uniqueness check: Ensure the uniqueness of key fields. Integrity check: Ensure the integrity of data records.

[0069] And based on the multi-modal industrial large model to batch identify and correct non-standard data. Such as classifying data, identifying different types of non-standard filling; predicting the reasonable value of data and correcting abnormal values; grouping similar data for batch processing.

[0070] Step 3.2, entity extraction and attribute extraction: Specifically, the hidden association between OPC UA information model modeling objects is mainly relied on the reasoning ability of the Chain of Thought enhanced large language model. After the above OPC UA information modeling and standardization processing, the data itself error is corrected, but the semantic relationship between each other objects is limited to the data quality has not been fully sorted out, so it is necessary to use the reasoning ability of the Chain of Thought enhanced large language model to carry out relationship mining and expansion.

[0071] The large language model itself has certain reasoning ability, but the Chain of Thought prompting method as a means of saving computing resources without training and fine-tuning the large language model can improve the reasoning ability of the large language model by breaking down complex problems into multiple-step sub-problems through prompt engineering. For example, the Chain of Thought enhanced GPT-3.5 model has significantly improved performance on mathematical problems, common sense reasoning, and symbolic reasoning datasets compared to the standard model.

[0072] Next, take the maintenance of the motor as an example to design the Chain of Thought prompting engineering:

[0073] Q: The motor overheats and cannot continue to run. What part is related?

[0074] A: Let's think step by step. Based on the collected data, list possible causes of the fault: high temperature due to cooling system failure, poor ventilation; abnormal current due to short circuit, winding damage... Logical reasoning: if the current temperature exceeds the normal range and the cooling system does not work properly, it can be inferred that the overheating is due to insufficient cooling.

[0075] A: Action taken: recommend checking the cooling liquid level, cleaning the radiator, etc.

[0076] Thus, the object relationship of motor overheating - cooling system - cooling liquid level / radiator is established. New reference relationship between each other is established.

[0077] Specifically, the completion of entity extraction and attribute extraction mainly relies on the natural language processing ability of large language models based on prompt engineering. Entity extraction in a knowledge graph refers to automatically identifying named entities from raw corpus. Entities are the most basic elements of a knowledge graph, and their extraction accuracy directly affects the quality of the subsequent knowledge base. Traditional entity extraction methods can be broadly divided into three categories: (1) rule-based methods, which are characterized by performing under limited semantic and text field conditions. Under defined rules, entity information is extracted, but this method relies heavily on expert experience and is difficult to adapt to new requirements of various changing data; (2) statistical learning-based entity extraction methods, which treat named entity recognition as a sequence labeling problem and use partially labeled or fully labeled semantic information for training; (3) deep learning-based methods, which have good fitting ability for complex nonlinear problems and can learn complex features. The large language model method based on prompt engineering belongs to method (3). Prompt engineering consists of three parts: instruction (Instruction), content (Content), and input text (InputText). The following four methods are used for prompt word design, and specific examples are as follows:

[0078] 1. Zero-shot (zero-shot learning):

[0079] First, Instruction (instruction): "According to the following OPC UA information model, extract the main body and attribute of the knowledge graph."

[0080] Content (content): NULL (null value).

[0081] InputText: MotorInstance: The specific motor instance, inherited from MotorType; Status: Instantiated StatusVariableType, representing the current status of the motor; Configuration: Instantiated ConfigurationVariableType, containing the specific configuration parameters of the motor.

[0082] Faults: FaultType; FaultDescription; FaultTime.

[0083] Troubleshooting: Solution; Steps; Processors.

[0084] AssociatedProducts: Associated product information.

[0085] The model provides the following answers: Key entity information includes Faults, MotorInstance, Manufacture Status / Configuration, and Troubleshooting; Key attribute information includes MotorInstance, Product No, Product Category, FaultDescription, Reporter, FaultType, Fault Cause, Craft No, CraftVersion, Resolution, Steps, and Processors.

[0086] 2. Few-Shot (Learning by Few Shots):

[0087] Instruction: Based on the following OPC UA information model, perform knowledge graph subject extraction and attribute extraction.

[0088] Content: MotorInstance-others (similar to MotorInstance).

[0089] InputText: Omitted (same as Zero-Shot).

[0090] Answer: Omitted.

[0091] 3. Role-Prompt:

[0092] Instruction: Based on the following OPC UA information model, perform knowledge graph subject extraction and attribute extraction.

[0093] Content: Suppose you are now a professional maintenance technician for xx wind turbine equipment.

[0094] InputText: Omitted (same as Zero-Shot).

[0095] Answer: Omitted.

[0096] 4. Chain-of-Thought:

[0097] Instruction: Based on the following OPC UA information model, perform knowledge graph subject extraction and attribute extraction.

[0098] Content: Let's think step by step.

[0099] InputText: Omitted (same as Zero-Shot).

[0100] Answer: Omitted.

[0101] Step 3.3, forming knowledge representation: Specifically, knowledge graph relation extraction and labeling are performed based on Relation Graph Convolutional Neural Network (R-GCN), but other graph convolutional networks or graph attention networks can also be used. R-GCN classifies relations into three categories: relations pointing to the node itself, relations pointing to external nodes, and relations pointing to the node itself. By aggregating the information vectors of these three types of relations, the corresponding probability distributions are obtained according to the task requirements through the ReLU function. Compared to GCN, R-GCN separates the edge information and the node's own information in the hidden layer, using Wr to represent and adjust the information of neighboring nodes on the edge, thus compensating for GCN's insufficient consideration of the information of nodes connected to the edge. Define the network G=(V,ε,R), where node vi ∈ V, edge (v i ,r,v j )∈ ε, where r∈ R represents the type of the graph network edge, i.e., the relationship type. The information of a specific node is calculated as follows:

[0102]

[0103] where represents the hidden layer information of the node v i in the l first layer, represents the hidden layer information of the node v i in the l+ 1st layer, represents the hidden layer information of the node v j in the l layer, and σ represents an activation function, represents the neighbor node of the node i with the r-type edge, is a normalization constant for a specific problem, which can be learned or pre-set, represents the weight matrix of the relationship r in the l layer, represents the feature transformation weight of the node itself in the l layer.

[0104] R-GCN needs to specify a transformation function for each type of edge. In order to reduce the computational overhead, the transformation function parameters of different types of relationships are shared, and a base function decomposition is used:

[0105]

[0106] represents a set of base matrices in the l layer, and there are B base matrices in total, is used to construct all relationship weights, is the linear combination coefficient of the relationship r to the base matrix .

[0107] For entity classification, a plurality of R-GCN layers are stacked, and a softmax function is passed after the output of the last layer. Ignoring the unlabeled nodes, the cross-entropy loss Loss1 is minimized for all labeled nodes as follows:

[0108]

[0109] where Y is the set of labeled nodes, K represents the output dimension, represents the output of the i-th labeled node at the k-th position, L represents the output layer, and t ik represents the true value of the label of the i-th labeled node at the k-th position.

[0110] On the basis of the foregoing implementation of entity extraction, attribute extraction, and relation extraction, a Neo4j graph database is imported to construct a knowledge graph (KG), KG={E,R,F}, where E represents a set of entities contained in the knowledge graph {e1, e2,..., e E}, R represents a set of relations between different entities {r1, r2,..., r R}, F represents a set of triples in the knowledge graph {f1, f2,..., f T}, each triple f is defined as (s, r, o), where s, r, and o represent a head entity, a relation, and a tail entity, respectively, and T represents the total number of triples.

[0111] Step 4, optimization and dynamic updating of the knowledge graph: the semantic relationship between node entities is optimized and updated using a graph convolution network and a long short-term memory network, knowledge merging and processing are completed, and the abnormal knowledge graph is dynamically updated. Specifically, it includes:

[0112] Step 4.1, knowledge merging and processing: the step 3 has completed the preliminary knowledge representation, but the knowledge graph at this time is still incomplete, resulting in slightly poor performance when executing downstream tasks. In order to avoid this situation, knowledge merging and processing are still needed for the knowledge graph, and a dynamic updating mechanism is constructed. Knowledge merging and processing mainly include link prediction, knowledge reasoning, and entity disambiguation, and the dynamic updating mechanism mainly relies on the LSTM long short-term memory network to update the parameter matrix of the graph convolution network to utilize the time series information.

[0113] Link prediction is based on the original knowledge graph—graph neural network encoder—decoder—general model prediction framework of knowledge graph link prediction. R-GCN uses a graph neural network to implement the link prediction task, and the process of updating the information of a single node in the displayed knowledge graph representation is shown, the information is collected from adjacent nodes, and the updated representation is aggregated to the node, which is iteratively updated through downstream tasks until the node vector reaches a fixed point. After knowledge graph representation learning, DisMult decoder is used to generate a possibility score for each potential edge in the graph, and finally the most possible edge prediction is selected. DisMult (Distributed Multiplicative Model) decoder is a conventional decoder model in the field of Knowledge Graph Embedding (KGE), which is not described here.

[0114] A triple (s, r, o) is defined, s and o represent two entities, the relation r and the diagonal matrix R r ∈R d×d The scoring function f of the triple (s, r, o) is calculated as follows:

[0115]

[0116] where e s is the mapped vector of entity s after the encoder composed of R-GCN, e o is the mapped vector of entity o after the encoder composed of R-GCN, and the superscript T denotes the transpose.

[0117] We use negative sampling training, for each positive sample (s, r, o), randomly destroy its head entity or tail entity, generate ω negative samples, the total training set T includes positive samples (y = 1) and negative samples (y = 0), y is the sample type identifier, and the cross entropy loss function Loss2 is designed as follows:

[0118]

[0119] where is a subset of edges E in the graph network G = (V, E, R), σ is a logical sigmoid function, which maps the score to the interval [0, 1], and represents the probability that the triple is true.

[0120] The score f(s, r, o) of DistMult determines the ranking order of candidate triples, and evaluation indicators (such as MRR, HITS@K) completely depend on this ranking result. Since the global reasoning of abnormal association needs to comprehensively evaluate the ranking quality, here we use the average reciprocal rank MRR:

[0121]

[0122] where S represents the triple set of the test set, i.e. F= {f1,f2,...,f T}, rank i represents the ranking of the i-th triple in the prediction result.

[0123] Knowledge reasoning is based on existing data, using specific methods to infer new relationships or identify incorrect information, to solve the problem of incomplete knowledge graph. General graph convolution network (GCN) can be used for knowledge reasoning, and the core is to learn the embedding representation of entities through graph convolution operation. GCN can capture the neighbor node information of entities, update the feature representation of nodes layer by layer, so as to effectively learn the complex relationship between entities.

[0124] GCN aggregates the information of neighbor nodes through graph convolution operation, and updates the feature representation of nodes. The basic graph convolution formula is:

[0125]

[0126] Where A represents the original adjacency matrix of the graph. This represents an adjacency matrix with self-joins, where I is the identity matrix and the degree matrix. , This indicates whether nodes i and j in the graph are adjacent; 1 indicates adjacent and 0 indicates otherwise. σ is the activation function, and H is the activation function. (l) It is the first l The node feature matrix of the layer, H (l+1) W is the node feature matrix of the (l+1)th layer. (l) It is the first l Layer weight matrix.

[0127] During the knowledge reasoning phase, GCN can use the learned entity embeddings to predict new triples. It calculates scores for all possible tail entities and selects the entity with the highest score as the prediction result.

[0128] Entity disambiguation is a challenging problem in knowledge graphs because real-world entities can appear in many different forms within the graph. Graph Convolutional Neural Networks (GCNs) can achieve entity disambiguation through the following process and formula:

[0129] After learning the embedding representation of each entity in the knowledge graph using GCN, contextual information aggregation is performed. For each entity reference in the text, it needs to be linked with the entity in the knowledge graph. This can be achieved by calculating the similarity between the context vector of the entity reference and the entity embedding vector in the knowledge graph. The similarity function can be cosine similarity, Euclidean distance, etc. Based on the similarity score, the most likely entity is selected as the disambiguation result of the entity reference. This is expressed by the following formula:

[0130] Score(e,m)=Similarity(H e H m )

[0131] Among them, H e H is the context vector referred to by entity e. m is the embedding vector of entity m in the knowledge graph, Score represents the calculated score, and Similarity is the similarity calculation function.

[0132] The loss function Loss3 during training can be expressed as:

[0133]

[0134] Where T is the set of triples in the training set, m' is the negative entity used for contrastive learning during training, and γ is a hyperparameter used to control the boundary of the score.

[0135] Step 4.2, Constructing the Dynamic Evolution of the Knowledge Graph: To overcome the problem that existing knowledge graphs, being mostly static, cannot update node structure and semantic information in a timely manner, a dynamic knowledge graph update mechanism will be introduced. Specifically, a graph convolutional neural network is used to capture the dependencies in the graph, and a Long Short-Term Memory (LSTM) network is used to learn and evolve the parameter matrix of the l-th layer of the GCN model at each time step t. This allows for the dynamic evolution of knowledge graphs constructed using knowledge embedding.

[0136] First, rewrite the expression for the convolutional neural network in terms of time t:

[0137]

[0138] In the formula, H (l) It is time t. l The node feature matrix of the layer, H (l+1) W is the feature matrix of the nodes in the (l+1)th layer at time t. (l) It is time t. l The weight matrix of the layer, It is the degree matrix at time t, for X t The input at time t is the initial feature matrix of the graph, which in this method represents the encoding matrix of each node in the graph.

[0139] The core of weight evolution is to use a Long Short-Term Memory (LSTM) network to update the weight matrix at time t based on current and historical information. 。 In typical graph convolutional networks It is constant, but in a dynamic knowledge graph, the graph structure at different times t may be different. In order to reflect the evolution of the graph structure over time, let It varies with the time step and the convolutional layer, that is:

[0140]

[0141] In the formula, LSTM represents Long Short-Term Memory Network, which is used to update... The specific update formula is as follows:

[0142]

[0143] in, This represents the parameter matrix of the l-th layer of the graphical convolutional network at time t-1. Represents the time t-1. l The node feature matrix of the layer, where σ represents the activation function.

[0144] Wf is the weight matrix of the forget gate, f t is the forget gate, determining how much old state to keep; W i is the weight matrix of the input gate, i t is the input gate, controlling the inflow of new information; W o is the weight matrix of the output gate, O t is the output gate, controlling the output of the current state; W c is the weight matrix of the candidate state, represents the candidate state, generating a new candidate value, c t represents the updated state; the final output parameter matrix , represents the Hadamard product, i.e. the element-wise multiplication of two matrices or vectors, and tanh represents the hyperbolic tangent function.

[0145] Using a long short-term memory network, the state at time t-1 is used to update , which can analyze the influence of atlas evolution at different times and realize dynamic updating of the knowledge graph.

[0146] Step 5: Knowledge graph reasoning application of wind turbine abnormal knowledge association: including intelligent retrieval of fault knowledge, identification of repetitive faults, role-based fault data analysis, effectiveness analysis of corrective measures, and predictive fault maintenance.

[0147] Intelligent retrieval based on fault knowledge base is mainly based on the entities, attributes, and associations designed in the knowledge graph. Through natural language processing methods, the triplets of knowledge are automatically extracted to form the core knowledge base. Intelligent retrieval has two forms. One is the question and answer form, which defines the template type of common questions. When the user asks a question, the question is structured through word segmentation tools, part-of-speech analysis, and syntax analysis to extract the structure, sentence pattern, and key entities of the question. The corresponding template of the question is identified. The best matching result is taken as the answer to the question and assembled into a sentence to return to the user. The other is the multi-keyword combination mode. The user inputs the fault or question keywords of interest. The fault or keywords are identified. If it is a term, the term is input into the graph to query related fault information. If it is not a term, only the keywords are combined to match to find the fault information implied in the fault description, thereby realizing intelligent fault information retrieval.

[0148] Repetitive failure refers to two failures although there are differences in the occurrence process and failure appearance, but the internal logic of the failure and the cause are consistent. This type of failure is an important concern of quality improvement. The fault knowledge base based on the knowledge graph can better complete the identification of repetitive failure, because the knowledge graph identifies the similarity between two fault entities through the entity alignment method, and realizes accurate query through the threshold value determined by the experiment.

[0149] Role-based fault analysis refers to the information in the fault knowledge base from different roles can improve product quality from different angles, and according to the actual business role, the fault knowledge of the design personnel, process personnel and inspection personnel is more concerned. Different roles have different weights on different fault types and attributes, and the potential value of the data is explored by using the knowledge graph, so as to help practitioners quickly find the problem root cause and solution.

[0150] Corrective action effectiveness analysis is an important part of fault management in the industrial field, and the fault entity is associated with the corresponding corrective action by applying the fault knowledge base based on the knowledge graph. The core idea is: taking a fault as an example, searching the number of repetitive failures in the past months since the fault occurs based on the repetitive failure identification method, and searching the number of repetitive failures in the future months since the fault occurs, dividing the two values to obtain the effectiveness coefficient, so as to evaluate the effectiveness of the corrective action.

[0151] Predictive maintenance is an important goal to ensure smooth operation of instruments and equipment and avoid major accident losses, and the abnormal knowledge association method of the wind turbine based on the knowledge graph can model the fault information of the device from the historical knowledge base, comprehensively consider the relevant semantic information, construct a nonlinear least squares function to fit the periodic life of the device, and realize the predictive maintenance of the device.

[0152] Referring to Figure 3 The embodiment of the application provides a kind of abnormal knowledge association device of wind turbine based on knowledge graph, including one or more processors, for realizing the knowledge association method of wind turbine based on knowledge graph in one of the above embodiments.

[0153] The implementation of the abnormal knowledge association device of wind turbine based on knowledge graph can be applied to any device with data processing capability, which can be a device such as a computer or the like. The device embodiment can be realized by software, or by hardware or a combination of software and hardware. Taking software implementation as an example, as a logical device, it is formed by reading the corresponding computer program instructions in the non-volatile memory into the memory and running by the processor of the device with data processing capability. From the hardware level, such as Figure 3As shown, it is a hardware structure diagram of any data processing capable device where the abnormal knowledge association device based on a knowledge graph of a wind turbine is located, in addition to Figure 3 In addition to the processor, the memory, the network interface, and the non-volatile memory shown, any data processing capable device where the device in the embodiment is located can also include other hardware according to the actual functions of the data processing capable device, and no further description is given.

[0154] The implementation process of the functions and roles of each unit in the above device is specifically described in the implementation process of the corresponding steps in the above method, and no further description is given here.

[0155] Each technical feature of the above-described embodiments can be combined arbitrarily, and to make the description concise, all possible combinations of the technical features in the above-described embodiments are not described, however, as long as the combinations of the technical features do not exist contradictory, it should be considered as the scope of the present application.

[0156] The embodiment of the present application also provides a readable storage medium having a program stored thereon, and the program is executed by a processor to implement the abnormal knowledge association method based on a knowledge graph of a wind turbine in the above-described embodiment.

[0157] The readable storage medium can be an internal storage unit of any data processing capable device, such as a hard disk or a memory. The readable storage medium can also be an external storage device, such as a plug-in hard disk, a smart media card (SMC), an SD card, a flash card, etc. Further, the readable storage medium can include both an internal storage unit of any data processing capable device and an external storage device. The readable storage medium is used to store the computer program and other programs and data required by the data processing capable device, and can also be used to temporarily store data that has been output or will be output.

[0158] The terms used in one or more embodiments of the present specification are only for the purpose of describing specific embodiments, and are not intended to limit one or more embodiments of the present specification. The singular forms "a", "an" and "the" used in one or more embodiments of the present specification and the appended claims are also intended to include the plural forms, unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used herein means and includes any or all possible combinations of one or more associated listed items.

[0159] It should be understood that, although the terms first, second, third, etc. can be used herein to describe various information, these terms are not intended to denote a particular order or hierarchy among the information. The terms are used only to distinguish one category of information from another. For example, a first information can be termed a second information, and, similarly, a second information can be termed a first information, without departing from the scope of the present description. As used herein, the term "if' can be interpreted to mean "when" or "upon" or "in response to determining" taking into account negation that can be associated with such interpretations. In other words, an "if" statement can be understood to mean "if or when" or "upon the occurrence of" or "in response to determining" at least one condition.

[0160] The foregoing description of the preferred embodiment of the present description has been presented for the purposes of illustration and description. It is not intended to be exhaustive or to limit the description to the precise form disclosed. Many modifications and variations are possible in light of this disclosure. It is intended that the scope of the description be limited not by this detailed description, but rather by the claims appended hereto.

Claims

1. A knowledge graph-based abnormal knowledge association method for a wind turbine, characterized in that, Comprise: Step 1, based on the operation and maintenance requirements of the wind power industry, distributed multi-source heterogeneous data and knowledge collection is carried out; Step 2, the OPC UA information model of the wind power equipment node is established, and the general attributes and reference types of the node are determined; Step 3, through the prompt word engineering, the industrial multi-modal industrial large model is used to standardize the description and identification of node data, complete entity extraction and attribute extraction, use graph convolution network to extract and represent the relationship, and import the graph database to construct the abnormal knowledge graph; Step 4, using graph convolution network and long short-term memory network to optimize and update the semantic relationship between node entities, complete knowledge merging and processing, dynamically update the abnormal knowledge graph, including: Step 4.1, knowledge merging and processing: based on the graph convolution neural network, the knowledge merging and processing process is completed, including link prediction, knowledge reasoning, entity disambiguation, and forming a standard knowledge representation; Step 4.2, dynamic evolution of knowledge graph: using graph convolution neural network to capture the dependency relationship of the graph, using long short-term memory network to learn and evolve the parameter matrix of different layers of the graph convolution network at each time, and using knowledge embedding to construct the dynamic evolution of the knowledge graph; Step 5, facing the abnormal knowledge application scene, based on the abnormal knowledge graph, the entity association is mined for knowledge reasoning.

2. The wind turbine unit abnormal knowledge association method based on a knowledge graph according to claim 1, characterized in that, The step 1 comprises: Step 1.1, multi-source knowledge data integration: the data sources include internal data sources and external data sources of wind power equipment operation and maintenance enterprises, the internal data sources include equipment real-time operation data, system operation and maintenance log, internal document data, the external data sources include industry standards and public data sets, academic literature and patents; Step 1.2, data knowledge preprocessing: according to the data quality, the feature engineering is constructed to replace the average value of abnormal data, missing data, data fitting and data rejection processing, and the knowledge source is reasonably tested to remove low-quality information sources.

3. The wind turbine unit abnormal knowledge association method based on a knowledge graph according to claim 1, characterized in that, The step 2 comprises: Step 2.1, define the OPC UA information model: the node types of the OPC UA information model include objects, object types, variables, variable types, data types, reference types, methods, and views, the general attributes of the OPC UA information model include node identification, node type, browse name, display name, description, write mask, and user write mask, the additional attributes of the OPC UA information model include whether it is abstract, whether it is symmetric, and reverse name; Step 2.2, build the OPC UA information model: the establishment of the OPC UA information model includes four steps, namely demand acquisition, definition type model, instantiation information model, and export XML and CSV documents.

4. The wind turbine unit abnormal knowledge association method based on a knowledge graph according to claim 1, characterized in that, The step 3 comprises: Step 3.1, standardize the OPC UA information model of wind power equipment: use the multi-modal industrial large model with industrial data fine-tuning to preprocess the OPC UA information model in step 2, clean up unreasonable structured information, and enhance the embedded reasoning ability in the form of thinking chain to expand the relationship between different individuals; Step 3.2, entity extraction and attribute extraction: based on various prompt word engineering methods, including zero-shot learning, small sample learning, role prompt, thinking chain, guiding the multi-modal industrial large model to normalize the OPC UA information model into the ontology construction under the attribute graph model, completing entity extraction and attribute extraction; Step 3.3, form knowledge representation: based on the relationship graph convolutional neural network to extract and represent the relationship on the preprocessed OPC UA information model data.

5. The wind turbine unit abnormal knowledge association method based on a knowledge graph according to claim 1, characterized in that, The link prediction in the step 4.1 is based on the original knowledge graph-graph neural network encoder-decoder-knowledge graph link prediction framework; the graph convolutional neural network uses the graph neural network to realize the link prediction task, explicitly represents the process of updating the information of a single node in the knowledge graph, collects information from adjacent nodes, aggregates and updates the representation of the node, and iteratively updates through the downstream task until the node vector reaches the fixed point. After abnormal knowledge graph learning, DisMult decoder is used to generate a possibility score for each potential edge in the graph, and finally the most possible edge prediction is selected; The knowledge reasoning in the step 4.1 includes: the graph convolutional neural network uses the learned entity embedding to predict new triples; By calculating the scores of all possible tail entities and selecting the entity with the highest score as the prediction result; The graph convolutional neural network in the step 4.1 realizes entity disambiguation through the following process: By calculating the similarity between the context vector of the entity reference and the entity embedding vector in the knowledge graph, each entity reference in the text is linked to the entity in the knowledge graph, and according to the similarity score, the most possible entity is selected as the disambiguation result of the entity reference.

6. The wind turbine unit abnormal knowledge association method based on a knowledge graph according to claim 1, characterized in that, The step 5 includes: Knowledge graph reasoning application of abnormal knowledge association of wind turbine: the knowledge graph reasoning application of abnormal knowledge association includes intelligent retrieval of fault knowledge, identification of repetitive faults, role-based fault data analysis, effectiveness analysis of corrective measures, and predictive fault maintenance.

7. A wind turbine unit abnormal knowledge association device based on a knowledge graph, characterized in that, One or more processors for implementing the knowledge graph-based wind turbine abnormal knowledge association method of any one of claims 1-6.

8. A readable storage medium, characterized by, A program is stored thereon, which is executed by the processor to implement the knowledge graph-based wind turbine abnormal knowledge association method of any one of claims 1-6.

Images