Wind turbine generator fault diagnosis method based on knowledge embedded heterogeneous graph

Abstract

The invention discloses a wind turbine generator fault diagnosis method based on knowledge embedded heterogeneous graph learning, which comprises the following steps: firstly, acquiring time sequence historical operation data of a multivariable sensor from a wind power plant, constructing a domain knowledge-based heterogeneous graph by using the processed multivariable time sequence data, and then carrying out heterogeneous graph representation learning; respectively inputting the domain knowledge heterogeneous graph into a heterogeneous interactive attention module to calculate an attention vector between nodes and a heterogeneous message passing module to calculate information between the nodes, and then weighting and aggregating the calculated attention vector and the information between the nodes from a source node to a target node; and obtaining the feature representation of each heterogeneous graph through end-to-end training. And finally, inputting the global graph representation into a full connection layer, outputting probabilities of different fault types by using a SoftMax classifier, and finally obtaining a fault type label of each state. According to the method, the heterogeneous graph embedded by the domain knowledge is constructed, so that fault features can be learned more effectively, and the fault diagnosis accuracy is improved.

Description

Technical Field

[0001] The present invention belongs to the field of wind turbine fault diagnosis, and relates to a wind turbine fault diagnosis method based on a knowledge-embedded heterogeneous graph. Background Art

[0002] Wind power generation, as a clean and renewable energy source, is an important part of the energy transition. The installed capacity of wind turbines in China has been continuously increasing, and the cumulative installed capacity has reached 440GW. However, due to the harsh operating environment of wind turbines and the poor accessibility of operation and maintenance, faults and low efficiency are inevitable during the long-term operation of wind turbines, which affect the power generation efficiency and even cause serious economic losses and safety hazards. Therefore, it is of great economic significance and social value to timely and accurately diagnose the faults of wind turbines, reduce the operation and maintenance costs, and ensure the stable and efficient operation of the units.

[0003] Currently, wind turbines have been widely installed with a Supervisory Control and Data Acquisition (SCADA) system. The SCADA system can monitor various operating parameters of wind turbines in real time, such as wind speed, rotational speed, power, temperature, voltage, current, etc. The massive and rich SCADA data provides data support for the intelligent fault diagnosis of wind turbines. The SCADA data is essentially multivariate time series data based on multiple sensors. Exploring the fault patterns in the signals of each sensor and the complex interactions between different components, and mining the spatio-temporal characteristics of multivariate time series data is an important and challenging task for the fault diagnosis of wind power systems.

[0004] In recent years, significant progress has been made in machine learning and deep learning methods based on SCADA data. Machine learning methods such as using decision trees and random forests are used to achieve fault identification and prediction. Deep learning methods such as convolutional neural networks (CNNs) and long short-term memory networks (LSTMs) improve the accuracy and robustness of diagnosis by automatically extracting complex features. However, traditional machine learning and deep learning methods mainly process Euclidean-structured data with neat arrangements and fail to fully utilize the correlation and dependence information between multi-sensor data of wind turbines. To further improve the accuracy and efficiency of fault diagnosis, it is particularly necessary to introduce graph learning methods. Graph data can effectively model and utilize the relationship and structural information between data, is suitable for processing multi-dimensional data of complex systems, and provides more comprehensive feature information for the fault diagnosis of wind turbines.

[0005] In addition, as a large-scale complex object composed of multiple subsystems, each subsystem of a wind turbine has a specific function, and the SCADA system includes multiple types of sensors. Each data varies in terms of frequency, amplitude, sensitivity, etc., and has its own unique feature distribution, that is, the physical structure-functional relationships of the subsystems and the characteristics of signal variables are highly heterogeneous. Therefore, embedding the domain knowledge of heterogeneous characteristics between different functional units of a wind turbine into graph data can not only enhance the feature expression ability of graph data to achieve more accurate fault diagnosis, but also provide ideas for graph interpretability research. Summary of the Invention

[0006] The purpose of the present invention is to provide a fault diagnosis method for wind turbines based on knowledge-embedded heterogeneous graphs, design a method for generating heterogeneous graphs with domain knowledge embedded, and fully exploit the heterogeneous information and coupling information in the heterogeneous graphs of wind turbines.

[0007] To achieve the above technical objectives, the technical solution of the present invention is as follows:

[0008] A fault diagnosis method for wind turbines based on knowledge-embedded heterogeneous graphs, the method includes the following contents:

[0009] Step S1: Obtain the historical data of multi-variable sensor time series under healthy and different fault states of the wind turbine from the wind turbine state monitoring and data acquisition (SCADA) system.

[0010] Step S2: According to domain knowledge, embed prior knowledge into the processed multi-variable sensor time series to construct graph structure data from the system level to the internal monitoring variable level. That is, each sensor is regarded as a node, the features of the time series slices of the sensor are used as the features of the node, and the correlation between nodes is used as the edge. Finally, a heterogeneous graph G=(V, X, ε, A, R) is obtained, where V and ε represent the multi-type node set and edge set respectively, X represents the node feature matrix, and each node v∈V and each edge e∈ε are respectively associated with their type mapping functions V→A and ε→R.

[0011] Step S3: Construct a heterogeneous graph representation learning module. Input the heterogeneous graph structure data obtained in Step 2 into the heterogeneous interaction attention module to calculate the attention vectors between nodes and the heterogeneous message passing module to calculate the information between nodes, and then weight-aggregate the calculated attention vectors and information from the source node to the target node. Finally, the feature representation of each heterogeneous graph is obtained through end-to-end training.

[0012] Step S4: Fault classification module. Input the learned heterogeneous graph feature representation into the fully connected network layer, and use the SoftMax classifier to output the probabilities of different fault types, and finally obtain the fault type labels of each state.

[0013] A further improvement of the technical solution of the present invention lies in: constructing domain knowledge-embedded heterogeneous graph data

[0014] For the construction of the domain knowledge-embedded heterogeneous graph, a top-down analysis method from the system level to the monitoring variable level is adopted to embed prior knowledge.

[0015] First, experts divide the wind turbine into 6 subsystems according to the domain knowledge of the physical topology structure of the wind turbine system: blade system, pitch system, bearing system, rotor system, grid system, and meteorological system. At the system level, a topological connection relationship guided by the spatial physical structure is established; at the monitoring variable level, each node represents a sensor monitoring variable. The sensor variables within the same system are homogeneous nodes, that is, isomorphic nodes, and the nodes between different systems belong to heterogeneous nodes.

[0016] Regarding the edge connection relationship in the heterogeneous graph, there may be multiple types of edge connection relationships between both isomorphic nodes and heterogeneous nodes. Therefore, an edge-level heterogeneous connection relationship is designed. For example, there may be both a non-linear relationship and a causal relationship between the wind speed of the nacelle weather station and the power on the generator side of the frequency converter. Because the change in wind speed is the direct cause of the change in the power on the generator side. An increase in wind speed will cause an increase in the rotational speed of the wind turbine rotor, thereby increasing the power output of the generator. At the same time, the two nodes exhibit different non-linear characteristics in different wind speed segments: at lower wind speeds, the power increases with the increase in wind speed, but the growth rate is slow. When the wind speed reaches a certain threshold, the growth rate of the power will accelerate until the rated power is reached. If the wind speed continues to increase beyond a certain threshold, the wind turbine will limit the increase in power by adjusting the blade angle and other measures for safety reasons. Therefore, when generating the heterogeneous graph, the edge connection relationships between isomorphic nodes and heterogeneous nodes are constructed separately.

[0017] In addition, in order to avoid the high computational cost caused by the high dimensionality and complexity of node features when directly using the original data slices as node features in complex heterogeneous graph learning, the node feature representation is refined when generating graph data. Only the first-order moment and the second-order moment of the original sequence slice, two statistical features, are used. These can not only reflect the dispersion degree and volatility of the sequence data but also retain the core information. The formulas for the first-order moment and the second-order moment are respectively:

[0018]

[0019] where, x i is the element value in the time series, and n is the total number of elements in the time series.

[0020] Therefore, based on the above analysis and the idea of constructing a heterogeneous graph, the heterogeneous graph is defined as: G = (V, X, ε, A, R), where V and ε represent the multi-type node set and edge set respectively. Each node v ∈ V and each edge e ∈ ε are respectively associated with their type mapping functions V → A and ε → R. X represents the node feature matrix, and each node feature x i = [μ, σ 2 ∈ X. For the edge e = (s, t) from the source node s to the target node t, its meta-relation (edge heterogeneous relation) can be expressed as which can be simplified and represented by .

[0021] A further improvement of the technical solution of the present invention lies in: the heterogeneous graph representation learning module

[0022] Heterogeneous graph representation learning aims to obtain meaningful vector representations for each node to perform downstream tasks such as fault classification. In order to fully consider the system heterogeneous-level types included in the heterogeneous graph and the heterogeneous attributes of their child nodes, as well as the heterogeneous graph topological structure composed of multi-type edges, a heterogeneous graph representation learning method for embedding domain knowledge of wind turbines into the heterogeneous graph is proposed, which mainly includes three components: meta-relation-aware heterogeneous interaction attention, heterogeneous information transfer from the source node, and heterogeneous information aggregation based on a specific target.

[0023] Heterogeneous interaction attention: The heterogeneous interaction attention mechanism is adopted to calculate the attention vector between the source node s and the target node t, and its elements are the normalized similarity scores between the target node and its adjacent source nodes. For the h-th attention head, the source node s is mapped to the K h (s) vector, and the target node t is mapped to Q h (t), and their dot product is calculated as the attention. The mapping formulas of K h (s) and Q h (t) are as follows:

[0024]

[0025] where and are linear mappings depending on the node type, used to capture the heterogeneity of the nodes. H l-1 [s] represents the input feature of node s at the (l - 1) layer. The dot product calculation between the vectors Q and K is based on the weight matrix of the edge type to capture different feature information between isomorphic nodes and heterogeneous nodes. The attention score of the h-th head of the edge e is calculated by the formula:

[0026]

[0027] where It is a prior tensor, aiming to distinguish the significance of each neighbor node to the target node and serve as an adaptive scaling of attention. Connecting h attention heads together can obtain the attention vector for each node pair. Finally, the attention vector ATT(s, e, t) is calculated through SoftMax normalization, and the calculation formula is as follows:

[0028]

[0029] Among them, N(t) is all neighbor nodes of the target node t, and H is the total number of heads of the multi-head attention.

[0030] Heterogeneous message passing: This process is parallel to the heterogeneous interaction attention, aiming to transfer the source node information to the target node. To obtain the message of the h-th head Similarly, types of source nodes are mapped into the message vector through and the matrix is used to merge the edge-level heterogeneity. The calculation formula is:

[0031]

[0032] Finally, all h heads are weighted to obtain the

[0033]

[0034] Heterogeneous information aggregation: After calculating the heterogeneous multi-head attention and messages, they are weighted and aggregated from the source nodes to the target nodes. Using the attention vector ATT(s, e, t) as the weight, the messages from the source nodes are weighted and aggregated to obtain the updated vector The calculation formula is:

[0035]

[0036] After aggregating all neighbor information from different feature distributions to the target node t, finally, the vector of the target node t is mapped back to its initial distribution based on a specific type, indexed by its node type For this purpose, the linear mapping is applied to the updated vector Then a residual connection is made as follows:

[0037]

[0038] In this way, the output of the l-th heterogeneous representation learning layer of the target node t is obtained which is also the input of the (l + 1)-th layer. Finally, the node representations of the entire graph are obtained through end-to-end training.

[0039] In addition, the method makes a stacked design of L layers for the heterogeneous graph representation learning module, aiming to enable each node to reach a large proportion of nodes of different types and relationships in the whole graph. Here, L is a hyperparameter, and its value should not be too large.

[0040] Due to the adoption of the above technical solution, the technical progress achieved by the present invention is as follows:

[0041] The heterogeneous graph generation method with domain knowledge embedding proposed by the present invention can extract more comprehensive fault information and features compared with the traditional measure-based graph construction method, and has interpretability.

[0042] The heterogeneous graph representation learning module proposed by the present invention, compared with the method of directly using a homogeneous graph network model for heterogeneous graph representation learning, comprehensively considers the heterogeneous types of the system, the heterogeneous attributes of its sub-nodes, and the heterogeneous graph topology structure composed of multi-type edges, fully excavates the heterogeneous information and spatial coupling information embedded in the proposed knowledge in the heterogeneous graph, can more effectively capture the spatio-temporal correlation between the wind power system sensors, improves the accuracy of wind turbine fault diagnosis, and enhances the interpretability of the wind turbine graph learning fault diagnosis. Description of the Drawings

[0043] Figure 1 is a schematic flowchart of the process of the present invention;

[0044] Figure 2 is a schematic diagram of the method for generating a heterogeneous graph of a wind turbine based on knowledge embedding of the present invention;

[0045] Figure 3 is a schematic diagram of the overall framework of the wind turbine fault diagnosis scheme based on the knowledge-embedded heterogeneous graph of the present invention. Detailed Embodiments

[0046] The present invention provides a wind turbine fault diagnosis method based on a knowledge-embedded heterogeneous graph. The core of this method is to construct graph structure data according to domain knowledge, and use a designed heterogeneous graph representation learning module to extract the representation features of the heterogeneous graph of the wind turbine to achieve the fault diagnosis of the wind turbine. In order to enable those skilled in the art to better understand the solution of the present invention, the present invention will be further described in detail below with reference to the drawings and specific embodiments.

[0047] As Figure 1 shown, it is a modular flowchart of a wind turbine fault diagnosis method based on a knowledge-embedded heterogeneous graph provided by a specific embodiment of the present invention. This method includes obtaining multivariate sensor time series from a wind farm, and then generating heterogeneous graph structure data through the embedding of domain expert knowledge. Then, heterogeneous graph representation learning is carried out to extract the representation features of the heterogeneous graph of the wind turbine. Finally, the obtained global graph representation is input into the FC layer, and the SoftMax classifier is used to output the probabilities of different fault types to obtain the fault type labels.

[0048] As shown Figure 2 in the figure, it is a schematic diagram of a method for generating a heterogeneous graph of wind turbines based on knowledge embedding provided by the present invention. The method includes the following contents:

[0049] Step S1: Design the topology structure of the graph. According to the domain knowledge of the physical topology structure of the wind turbine system, the wind turbine is divided into 6 subsystems: blade system, pitch system, bearing system, rotor system, power grid system, and meteorological system. At the system level, a topological connection relationship guided by the spatial physical structure is established; at the monitoring variable level, each node represents a sensor monitoring variable. The sensor variables within the same system are homogeneous nodes, that is, isomorphic nodes, and the nodes between different systems belong to heterogeneous nodes.

[0050] Step S2: Design the edge-level connection relationship. The edge connection relationships between homogeneous nodes and heterogeneous nodes are constructed respectively. For the edge connection relationships in the heterogeneous graph, there may be multiple types of edge connection relationships between both homogeneous nodes and heterogeneous nodes. For example, there may be both a non-linear relationship and a causal relationship between the source node and the target node.

[0051] Step S3: Design the node features. Two statistical features, namely the first moment and the second moment of the original sequence slice, are used as node features. The formulas for the first moment and the second moment are respectively:

[0052]

[0053] where x i is the element value in the time series, and n is the total number of elements in the time series.

[0054] Step S4: The finally constructed heterogeneous graph is defined as: G=(V, X, ε, A, R), where V and ε respectively represent the multi-type node set and the edge set. Each node v∈V and each edge e∈ε are respectively associated with their type mapping functions V→A and ε→R. X represents the node feature matrix, and each node feature x i =[μ, σ 2 ∈X. For the edge e=(s, t) from the source node s to the target node t, its meta-relationship (edge heterogeneous relationship) is expressed as Simplified as represented by.

[0055] As Figure 3 shown in the figure, it is a schematic diagram of the overall framework of a specific implementation manner of the present invention, including the following contents:

[0056] Step S1: Obtain the historical data of multi-variable sensor time series under the healthy and different fault states of multiple wind turbines from the wind farm state monitoring and data acquisition (SCADA) system.

[0057] Step S2: According to domain knowledge, the processed multi-variable sensor time series is used to construct graph-structured data by embedding prior knowledge from the system level to the internal monitoring variable level. That is, each sensor is regarded as a node, the features of the time series slices of the sensor are used as the features of the node, and the correlation between nodes is used as the edge. Finally, a heterogeneous graph G = (V, X, ε, A, R) is obtained, where V and ε represent the multi-type node set and edge set respectively, X represents the node feature matrix, and each node v ∈ V and each edge e ∈ ε are respectively associated with their type mapping functions V → A and ε → R.

[0058] Step S3: Construct a heterogeneous graph representation learning module. The heterogeneous graph-structured data obtained in Step 2 is respectively input into the heterogeneous interaction attention module to calculate the attention vectors between nodes and the heterogeneous message passing module to calculate the information between nodes, and then the calculated attention vectors and information are weighted and aggregated from the source node to the target node. Finally, the feature representation of each heterogeneous graph is obtained through end-to-end training.

[0059] Step S4: Fault classification module. The learned heterogeneous graph feature representation is input into the fully connected network layer, and the SoftMax classifier is used to output the probabilities of different fault types. Finally, the fault type labels of each state are obtained.

[0060] The data source of this embodiment is the real-world direct-drive wind turbine operation data provided by a domestic wind power company, including two types of data: normal and blade cracking fault samples. Each sample is a csv file with a data sampling interval of 10 minutes, and each sample has approximately 400 time series data. The samples are all labeled data. This wind turbine contains 4620 samples, among which 2439 are normal samples and 2181 are fault samples. The original data set collected more than 75 different variables, and these original variables contain a lot of redundant information. Some variables have nothing to do with blade cracking or even the health status of the wind turbine. Therefore, based on the mechanism knowledge of the wind turbine, 25 effective variables related to blade cracking faults are selected by expert experience, including hub speed, vibration, wind speed, wind direction, power, current, and temperatures of various components. According to the knowledge-embedding-based heterogeneous graph generation method proposed by the present invention, the variables are classified into system level and monitoring variable level according to prior knowledge, as shown in Table 2:

[0061] Table 1 Information Table of System and Monitoring Variables

[0062]

[0063] The node features and edge features are designed as described in Step S2 of the claims. Then, after obtaining the global feature representation through graph representation learning, fault classification is carried out, and the fault classification results are shown in Table 2.

[0064] Table 2 compares the diagnostic results obtained by the embodiment with those of other related methods. The embodiment surpasses the fault diagnosis performance of other related methods in all indicators. ChebyNet, GraphSage, and CausalGNN are unable to capture diverse fault information from different representation spaces, resulting in incomplete fault information extraction. Furthermore, each isomorphic graph is generated through a homogeneous graph representation learning module, which essentially uses the concept of "homogenization" and cannot express the complex interactions between different types of nodes and edges. This loss of critical heterogeneous information leads to poor diagnostic results for specific categories.

[0065] The present invention has significantly improved the effect, and its superior performance can be attributed to the advanced mechanism of embedding domain knowledge into a multi-level heterogeneous graph structure and a heterogeneous graph learning module. On the one hand, the embedding of domain knowledge can guide the heterogeneous graph to model multi-level information at the system level and the sensor data level, allowing the model to mine the complex interaction relationships and heterogeneous relationships between different subsystems. On the other hand, the heterogeneous graph learning module directly extracts, transmits and aggregates information on the heterogeneous graph, can simultaneously consider different types of nodes and edges, and flexibly assign different weights to heterogeneous edges, which is more conducive to mining the complex coupling and heterogeneous characteristics of SCADA data.

[0066] Table 2 Comparison of the fault diagnosis test results of the present invention and related methods

[0067]

[0068] The embodiments described above are merely descriptions of preferred implementations of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should fall within the scope of protection determined by the claims of the present invention.

Claims

1. A fault diagnosis method for wind turbine generators based on knowledge-embedded heterogeneous graphs, characterized in that: The method includes the following steps: Step S1: Obtain the historical data of multivariate sensor time series under the healthy and different fault states of the wind turbine from the wind turbine condition monitoring and data acquisition (SCADA) system. Step S2: Domain knowledge is embedded into the heterogeneous graph generation module. According to domain knowledge, the processed historical data of multivariate sensor time series is used to construct graph structure data by embedding prior knowledge from the system level to the internal monitoring variable level. That is, each sensor is regarded as a node, the features of the time series slices of the sensor are used as the features of the node, and the correlation between nodes is used as the edge. Finally, the heterogeneous graph G = (V, X, ε, A, R) is obtained, where V and ε represent the multi-type node set and edge set respectively, X represents the node feature matrix, and each node v ∈ V and each edge e ∈ ε are respectively associated with their type mapping functions V → A and ε → R. Specifically, constructing the heterogeneous graph structure data based on domain knowledge in Step S2 includes: Step S2-1: Design the graph topology. According to the domain knowledge of the physical topology structure of the wind turbine system, the wind turbine is divided into 6 subsystems: blade system, pitch system, bearing system, rotor system, grid system, and meteorological system. At the system level, a topological connection relationship guided by the spatial physical structure is established; at the monitoring variable level, each node represents a sensor monitoring variable, and the sensor variables within the same system are of the same type of nodes, that is, isomorphic nodes, and the nodes between different systems belong to heterogeneous nodes. Step S2-2: Design the edge-level connection relationship. The edge connection relationships between isomorphic nodes and heterogeneous nodes are constructed respectively. For the edge connection relationships in the heterogeneous graph, there may be various types of edge connection relationships between isomorphic nodes and heterogeneous nodes. For example, there may be both non-linear relationships and causal relationships between the source node and the target node. Step S2-3: Design the node features. Use the two statistical features of the first moment and the second moment of the original sequence slice as the node features. The formulas for the first moment and the second moment are respectively: where x i is the element value in the time series, and n is the total number of elements in the time series. Step S2-4: The finally constructed heterogeneous graph is defined as: G = (V, X, ε, A, R), where V and ε represent the multi-type node set and edge set respectively. Each node v ∈ V and each edge e ∈ ε are respectively associated with their type mapping functions V → A and ε → R. X represents the node feature matrix, and each node feature x i = [μ, σ 2 ∈ X. For an edge e = (s, t) from the source node s to the target node t, its meta-relation (edge heterogeneous relation) is expressed as Simplified using Denoted by Step S3: Construct the heterogeneous graph representation learning module. The heterogeneous graph structure data obtained in Step 2 is respectively input into the heterogeneous interaction attention module to calculate the attention vector between nodes and the heterogeneous message passing module to calculate the information between nodes, and then the calculated attention vector and the information between nodes are weighted and aggregated from the source node to the target node. Finally, the feature representation of each heterogeneous graph is obtained through end-to-end training. The heterogeneous graph representation learning method described in Step S3: Step S3-1: Heterogeneous interaction attention. Use the heterogeneous interaction attention mechanism to calculate the attention vector between the source node s and the target node t, and its elements are the normalized similarity scores between the target node and its adjacent source nodes. For the h-th attention head, map the source node s to the K h (s) vector, map the target node t to Q h (t), and calculate their dot product as the attention. The mapping formulas for K h (s) and Q h (t) are as follows: in and It is a linear mapping that depends on the node type and is used to capture the heterogeneity of nodes. l-1 [s] represents the input feature of the (l-1)th layer node s. The dot product calculation between vectors Q and K is based on the edge type The weight matrix To capture the different feature information between homogeneous nodes and heterogeneous nodes. The attention score of the h-th head of edge e The calculation formula is: Among them, is a prior tensor, aiming to distinguish the significance of each neighbor node to the target node and serve as an adaptive scaling of attention. Connecting h attention heads together can obtain the attention vector for each node pair, and finally, the attention vector ATT(s, e, t) is calculated through SoftMax normalization. The calculation formula is as follows: where N(t) is all the neighbor nodes of the target node t, and H is the total number of heads of the multi-head attention. Step S3-2: Heterogeneous message passing. This process is parallel to heterogeneous interaction, aiming to transfer source node information to the target node. To obtain the message of the h-th head Similarly, source nodes of this type are mapped into the message vector, and the matrix is used to merge edge-level heterogeneity. The calculation formula is: Finally, all h heads are weighted to obtain Step S3-3: Heterogeneous information aggregation. After calculating the heterogeneous multi-head attention and messages, they are weighted and aggregated from the source nodes to the target node. Using the attention vector ATT(s, e, t) as the weight, the messages from the source nodes are weighted and aggregated to obtain the update vector The calculation formula is as follows: After aggregating all neighbor information from different feature distributions to the target node t, finally map the vector of the target node t back to its initial distribution based on a specific type, indexed by its node type For this purpose, apply the linear mapping to the updated vector Then perform a residual connection as follows: Step S3-4: Obtain the output of the l-th heterogeneous representation learning layer of the target node t That is, the input of the (l + 1)-th layer. Finally, the node representations of the entire graph are obtained through end-to-end training. Step S4: Fault classification module. Input the learned heterogeneous graph feature representation into the fully connected network layer, and use the SoftMax classifier to output the probabilities of different fault types, and finally obtain the fault type labels of each state.

2. The method for fault diagnosis of a wind turbine based on a knowledge-embedded heterogeneous graph according to claim 1, wherein The L coefficient of the stacked design of the L layers of the heterogeneous graph representation learning module is automatically optimized during the network training process.

Images