A wind turbine fault diagnosis method based on multi-view graph neural network

Abstract

The application discloses a wind turbine fault diagnosis method based on a multi-view graph neural network. First, multi-variable sensor time series historical data are obtained from a wind farm, and processed multi-variable sensor time series data are constructed into multi-view graph structure data from different perspectives. Then, graph representation learning is performed on each kind of graph structure data, a plurality of independent and parallel graph convolutional neural networks are inputted for feature extraction, a readout layer is used to aggregate node features after a graph pooling layer, and multi-graph feature representation is obtained. Then, the multi-graph feature representation is inputted into a multi-graph interactive learning fusion module to obtain multi-graph shared representation, and the multi-graph shared representation is spliced to obtain fused global graph representation. Finally, the global graph representation is inputted into an FC layer, and a softmax classifier is used to output probabilities of different fault types, and finally, fault type labels of each state are obtained. The application can learn more effective fault features through multi-graph interactive learning fusion, and improve fault diagnosis precision.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of wind turbine fault diagnosis, and relates to a wind turbine fault diagnosis method based on a multi-view graph neural network. BACKGROUND

[0002] Wind power is one of the most promising clean energies at present, and is also one of China's strategic emerging industries, and has developed in a large scale, high speed and large capacity in recent years. Due to the particularity of the operating environment of wind turbines, combined with the complexity of the internal structure and control strategy, the problems of high failure rate, low operating efficiency and poor reliability of the wind turbines in service are increasingly prominent. Once a component fails, it may trigger a chain reaction, causing the entire system to shut down or causing a major accident, resulting in serious economic losses. Therefore, accurate and timely fault diagnosis of wind power and early judgment of fault occurrence and fault type are of great significance and social value for reducing operation and maintenance costs and improving power generation efficiency. At present, wind turbine units have been widely installed with supervisory control and data acquisition (SCADA) systems, and SCADA data are multivariate time series data containing various sensor data. Due to various fault modes in the sensor signals and the complex interaction between different components, exploring the spatial interaction of sensors and mining the spatio-temporal features of multivariate time series data is an important and challenging task for wind power system fault diagnosis.

[0003] In recent years, great progress has been made in deep learning-based fault diagnosis methods, such as the wide application of convolution operations in SCADA data modeling. However, more and more researches in wind power systems consider the relationship and interdependence of data and represent the data in the form of graphs. Such irregular data in non-Euclidean space poses a great challenge to existing deep learning-based methods, making some important operations (such as convolution) easily applied in Euclidean space, but difficult to model graph data in non-Euclidean space. Recently, graph neural networks (GNN) have been used to model and analyze graph data as emerging neural networks.

[0004] However, due to the variety of wind turbine fault types, using a single graph construction method to extract fault features may not be able to capture all relationships comprehensively. The method of constructing multiple graphs can extract fault features from more perspectives and utilize the complementarity between graphs to realize interactive learning and fusion between multiple graph features, greatly improving the accuracy of fault diagnosis. SUMMARY

[0005] The application aims to provide a wind turbine fault diagnosis method based on a multi-view graph neural network.

[0006] To achieve the above technical purposes, the technical scheme of the application is as follows:

[0007] A wind turbine fault diagnosis method based on a multi-view graph neural network, which comprises the following contents:

[0008] Step S1: Obtain multivariate sensor time series historical data under health and different fault states of multiple wind turbines from a wind farm state monitoring and data acquisition (SCADA) system.

[0009] Step S2: Process the multivariate sensor time series data from different perspectives to construct multi-view graph structure data. That is, each sensor is regarded as a node, the time series slice of the sensor is regarded as the node feature, the correlation between two sensors is calculated as an edge, and distance graph neighbor graph functional graph causal graph form a multi-view graph structure data set

[0010] Step S3: Construct a graph representation learning module. The four kinds of graph structure data obtained in step 2 are respectively input into the corresponding graph convolutional neural network (GCN) for feature learning, and then pass through the graph pooling layer (Pooling), and finally use the readout layer (Readout) to aggregate the node features to obtain the graph feature representation corresponding to each graph structure data, respectively

[0011] Step S4: Construct a multi-graph interaction learning fusion module. The multi-graph features input into the interaction learning unit, define the interaction learning weight coefficient matrix α to assign different weights to the multi-graph features, utilize the correlation and complementarity between the multi-graph features, and finally learn the shared representation between the multi-graph features The formula is as follows

[0012]

[0013]

[0014] wherein, α AB = α BA , α AC = α CA , α AD = α DA , α BC = αCB ,α BD =α DB ,α CD =α DC ;α AA ,α BB ,α CC ,α DD respectively represent the autocorrelation measurement coefficients of distance graph, neighborhood graph, function graph and causal graph;α AB ,α BA represent the correlation measurement coefficients between distance graph and neighborhood graph;α AC ,α CA represent the correlation measurement coefficients between distance graph and function graph;α AD ,α DA represent the correlation measurement coefficients between distance graph and causal graph;α BC ,α CB represent the correlation measurement coefficients between neighborhood graph and function graph;α BD ,α DB represent the correlation measurement coefficients between neighborhood graph and causal graph;α CD ,α DC represent the correlation measurement coefficients between function graph and causal graph; these coefficients are automatically optimized in the network training process. Finally, the global graph representation f' is obtained by concatenation operation, and the formula is as follows.

[0015]

[0016] wherein, CONCAT represents concatenation operation.

[0017] Step S5: fault classification module. The fused global graph representation f' is input into a fully connected (FC) layer, and a softmax classifier is used to output the probability of different fault types, so as to finally obtain the fault type label of each state.

[0018] The further improvement of the technical scheme of the application is that the multi-view graph structure data is constructed, comprising:

[0019] The distance graph is constructed by finding the k' nearest neighbors of the node in other nodes of the sample through a Gaussian kernel weight function, calculating the edge weight value between the two nodes, and the formula is as follows

[0020]

[0021] wherein, represents the edge weight value between node x i and node x j , and ζ is the bandwidth of the Gaussian kernel.

[0022] The neighborhood graph is constructed​ is defined according to the actual spatial position of each sensor component of the wind turbine, if two sensors are adjacent, the edge weight between the two nodes is defined as 1; 0 if not adjacent, the formula is as follows

[0023]

[0024] The constructed functional graph The edge weight of the functional graph represents the functional correlation of each two sensors. The edge weight is determined by the Pearson correlation coefficient between the two nodes, and the calculation formula is as follows

[0025]

[0026] Wherein, x i,t and x j,t are the attribute values of node i and node j in period t, is the mean value of all nodes in period t attribute, and T represents the total length of time series.

[0027] The constructed causal graph is used to extract the causal relationship between two nodes, and the edge weight between two nodes is calculated by the transfer entropy value, and the formula is as follows.

[0028]

[0029] Wherein, represents the causal relationship of sensor variables i to j in the causal graph, P(x i+1 ,x i ,x j ) represents the joint probability of x i+1 ,x i ,x j ; P(x i+1 |x i ) represents the conditional probability of x i under the condition that x i+1 is known at i moment; P(x i ,x j ) represents the joint probability of x i and x j .

[0030] The further improvement of the technical scheme of the application is that the readout layer in step S3 considers the global mean value characteristics and local maximum value characteristics of the multi-view graph neural network, and the calculation formula is as follows

[0031]

[0032] Wherein, CONCAT represents mean pool and maximum pool The series operation of the two, N is the number of nodes, The feature representation of the i-th node, and finally the graph representation F of each graph is obtained.

[0033] Due to the adoption of the above technical solutions, the technical progress achieved by the present application is:

[0034] The multi-view graph structure data generation module provided by the present application can extract more comprehensive fault information and features compared to a single composition method, and has interpretability.

[0035] The multi-graph interactive learning fusion module provided by the present application realizes adaptive interaction between multi-graph features compared to a direct concatenation of multi-graph features, and the interactive learning unit can more effectively capture the spatio-temporal correlation between wind power system sensors by utilizing the correlation and complementarity between multi-graph features, thereby improving the accuracy of wind turbine fault diagnosis and facilitating timely processing and maintenance of wind turbines by staff. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 is a flowchart of the present application;

[0037] Figure 2 is a schematic diagram of the overall framework structure of the present application;

[0038] Figure 3 is a schematic diagram of the multi-graph interactive learning fusion module of the present application. DETAILED DESCRIPTION

[0039] The present application provides a wind turbine fault diagnosis method based on a multi-view graph neural network, the core of which is to construct multi-view graph structure data and realize wind turbine fault diagnosis by a multi-graph interactive learning fusion method. In order to enable personnel in the technical field to better understand the present application scheme, the present application will be further described in detail below in conjunction with the drawings and specific embodiments.

[0040] As shown in Figure 1 is a modular flowchart of a wind turbine fault diagnosis method based on a multi-view graph neural network provided by a specific embodiment of the present application. The method includes obtaining multi-variable sensor time series from a wind farm, then constructing distance graphs, neighborhood graphs, function graphs, and causal graphs through a multi-view graph structure data generation module to form a multi-graph set. Then, graph representation learning is performed on each kind of graph structure data, and then multi-graph features are interactively learned and fused. Finally, the fused global graph representation is input into an FC layer, and a softmax classifier outputs the probability of different fault types, and finally the fault type label is obtained.

[0041] As shown in Figure 2The diagram shown illustrates the specific framework of a wind turbine fault diagnosis method based on a multi-view graph neural network, according to a specific embodiment of the present invention. The method includes the following:

[0042] Step S1: Obtain multivariable sensor time series historical data of multiple wind turbines under different fault conditions from the wind farm condition monitoring and data acquisition (SCADA) system.

[0043] Step S2: Construct a multi-view graph structure from different perspectives using the processed multivariate sensor time series data. This involves treating each sensor as a node, using the sensor's time series slices as features of that node, calculating the correlation between each pair of sensors as edges, and constructing distance graphs accordingly. Neighborhood graph Functional diagram Cause and effect diagram Forming a multi-view graph structured data set

[0044] The constructed distance map The k′ nearest neighbors of a given node are found among other nodes in the sample using the Gaussian kernel weight function, and the edge weight between the two nodes is calculated using the following formula:

[0045]

[0046] in, Represents node x i and node x j The edge weights between them are given by ζ, which is the bandwidth of the Gaussian kernel.

[0047] The construction of the neighborhood graph It is defined based on the actual spatial positions of the various sensor components of the wind turbine. If two sensors are adjacent, the edge weight between these two nodes is... Defined as 1; 0 if not adjacent, the formula is as follows.

[0048]

[0049] The construction function diagram The edge weights represent the functional correlation between any two sensors. The edge weights are determined by the Pearson correlation coefficient between the two nodes, calculated as follows:

[0050]

[0051] Where, x i,t and x j,t These are the attribute values ​​of node i and node j within time period t. Let t be the mean of the attributes of all nodes in time period t, and T represent the total length of the time series.

[0052] the constructed causal graph For extracting the causal relationship between two nodes, the edge weight value between the two nodes is calculated by the transfer entropy value, and the formula is as follows.

[0053]

[0054] wherein, represents the causal relationship between sensor variables i and j in the causal graph, P(x i+1 ,x i ,x j ) represents the joint probability of x i+1 ,x i ,x j ; P(x i+1 |x i ) represents the conditional probability of x i under the condition that x i+1 is known at i; P(x i ,x j ) represents the joint probability of x i and x j .

[0055] Step S3: Construct a graph representation learning module. The four kinds of graph structure data obtained in step 2 are respectively input into the corresponding graph convolutional neural network (GCN) for feature learning, and then pass through the graph pooling layer (Pooling), and finally use the readout layer (Readout) to aggregate the node features to obtain the graph feature representation corresponding to each graph structure data, respectively

[0056] The readout layer considers the global mean value characteristics and local maximum value characteristics of the multi-view graph neural network, and the calculation formula is as follows

[0057]

[0058] wherein, CONCAT represents the concatenation operation of the mean pool and the maximum pool , N is the number of nodes, is the feature representation of the i th node, and finally the graph representation F of each graph is obtained.

[0059] As shown in Figure 3 , it is a multi-graph interaction learning fusion module schematic diagram, which is characterized by comprising the following contents:

[0060] Step S4: Construct a multi-graph interaction learning fusion module. The multi-graph features Input interaction learning unit, define interaction learning weight coefficient matrix α to assign different weights to multi-graph features, use the correlation and complementarity between multi-graph features, and finally learn the shared representation between multi-graph features The formula is as follows

[0061]

[0062]

[0063] wherein, α AB = α BA , α AC = α CA , α AD = α DA , α BC = α CB , α BD = α DB , α CD = α DC ; α AA , α BB , α CC , α DD respectively represent the autocorrelation measurement coefficients of distance graph, neighborhood graph, function graph and causal graph; α AB , α BA represent the correlation measurement coefficients between distance graph and neighborhood graph; α AC , α CA represent the correlation measurement coefficients between distance graph and function graph; α AD , α DA represent the correlation measurement coefficients between distance graph and causal graph; α BC , α CB represent the correlation measurement coefficients between neighborhood graph and function graph; α BD , α DB represent the correlation measurement coefficients between neighborhood graph and causal graph; α CD , α DC represent the correlation measurement coefficients between function graph and causal graph; these coefficients are automatically optimized during network training. Finally, the global graph representation f′ is obtained by concatenation operation, and the formula is as follows.

[0064]

[0065] wherein, CONCAT represents concatenation operation.

[0066] Step S5: fault classification module. The fused global graph representation f′ is input into the fully connected (FC) layer, and the softmax classifier is used to output the probability of different fault types, and finally the fault type label of each state is obtained.

[0067] The above-described embodiments are merely intended to describe the preferred embodiments of the present application, and are not intended to limit the scope of the present application. Various changes and modifications made by those skilled in the art to the present application without departing from the spirit of the present application should fall within the scope of the present application defined by the claims.

Claims

1. A method for fault diagnosis of wind turbine generators based on multi-view graph neural networks, characterized in that: The method includes the following steps: Step S1: Obtain multivariable sensor time series historical data on the health and different fault states of multiple wind turbine units from the wind farm condition monitoring and data acquisition system; Step S2: Construct multi-view graph structured data from different perspectives using the processed multivariate sensor time series data; that is, treat each sensor as a node, the time series slices of the sensor as the features of that node, calculate the correlation between any two sensors as edges, and construct distance graphs accordingly. Neighborhood graph Functional diagram Cause and effect diagram Forming a multi-view graph structured data set Step S3: Construct the graph representation learning module; The four types of graph structure data obtained in step S2 are input into the corresponding graph convolutional neural networks for feature learning, then processed through graph pooling layers, and finally aggregated using readout layers to obtain the graph feature representations corresponding to each graph structure data. Step S4: Construct a multi-image interactive learning fusion module to integrate the multi-image features obtained in Step S3. The input interactive learning unit defines an interactive learning weight coefficient matrix α to assign different weights to multiple graph features. By leveraging the correlation and complementarity among the multiple graph features, a shared representation among the multiple graph features is ultimately learned. The formula is as follows Where, α AB =α BA α AC =α CA α AD =α DA α BC =α CB α BD =α DB α cD =α DC ;α AA ,α BB ,α CC ,α DD These represent the autocorrelation coefficients of distance graphs, neighborhood graphs, function graphs, and causal graphs, respectively; α AB ,α BA α represents the correlation coefficient between the distance graph and the neighborhood graph. AC ,α CA α represents the correlation coefficient between the distance map and the function map. AD ,α DA α represents the correlation coefficient between the distance graph and the cause-effect graph. BC ,α CB α represents the correlation coefficient between the neighborhood graph and the function graph. BD ,α DB α represents the correlation coefficient between the neighborhood graph and the causal graph. CD ,α DC The correlation coefficient between the function graph and the cause-effect graph is represented by the following formula; finally, the global graph representation f′ is obtained through concatenation operations. Where CONCAT represents concatenation operation; Step S5: The fault classification module inputs the fused global graph representation f′ into the fully connected layer and uses a softmax classifier to output the probability of different fault types, finally obtaining the fault type label for each state.

2. The wind turbine fault diagnosis method based on a multi-view graph neural network as described in claim 1, characterized in that... Step S2, which involves constructing the multi-view graph structure data, specifically includes: The constructed distance map The k′ nearest neighbors of a given node are found among other nodes in the sample using the Gaussian kernel weight function, and the edge weight between the two nodes is calculated using the following formula: in, Represents node x i and node x j The edge weights between them, where ζ is the bandwidth of the Gaussian kernel; The construction of the neighborhood graph It is defined based on the actual spatial positions of the various sensor components of the wind turbine. If two sensors are adjacent, the edge weight between these two nodes is... Defined as 1; 0 if not adjacent, the formula is as follows. The construction function diagram The edge weights represent the functional correlation between any two sensors. The edge weights are determined by the Pearson correlation coefficient between the two nodes, and the calculation formula is as follows: Where, x i,t and x j,t These are the attribute values ​​of node i and node j within time period t. Let be the mean of the attributes of all nodes in time period t, where T represents the total length of the time series; The construction of causal graphs This is used to extract the causal relationship between two nodes. The edge weight between the two nodes is calculated using the propagation entropy value, as shown in the following formula. in, P(x) represents the causal relationship from sensor variable i to j in a causal graph. i+1 ,x i ,x j ) represents x i+1 ,x i ,x j The joint probability; P(x) i+1 |x i () indicates that x is known at time i. i In the case of x i+1 The conditional probability; P(x) i ,x j ) represents x i and x j The joint probability.

3. The wind turbine fault diagnosis method based on a multi-view graph neural network as described in claim 1, characterized in that: The readout layer described in step S3, considering the global mean and local maximum characteristics of the multi-view graph neural network, is calculated using the following formula: Where CONCAT represents the mean pool and the maximum pool The concatenation operation, where N is the number of nodes. Let F be the feature representation of the i-th node, and finally obtain the graph representation F of each graph.

4. The wind turbine fault diagnosis method based on a multi-view graph neural network as described in claim 1, characterized in that: α AC ,α CA ,α AB ,α BA ,α AD ,α DA ,α BC ,α CB ,α CD ,α DC ,α BD ,α DB These coefficients are automatically optimized during network training.

Images