Wind turbine generator time-varying working condition fault diagnosis method based on space-time multi-scale graph convolutional neural network

Abstract

The invention discloses a wind turbine generator time-varying working condition fault diagnosis method based on a space-time multi-scale graph convolutional neural network, and belongs to the technical field of equipment fault diagnosis, and the method comprises the following steps: data collection, data preprocessing, space-time feature extraction and fusion, and diagnosis result output. Aiming at the obvious time-varying working condition characteristic of SCADA data, a predefined and adaptive fused time-space diagram is constructed between any two time steps, and the dynamic dependency relationship between different nodes under different time steps is mined. Meanwhile, an extended causal space-time convolution layer is introduced, and potential time features are captured. Classification diagnosis of complex fault modes under the time-varying working condition is achieved, and a new technical solution is provided for the field of wind turbine generator fault diagnosis.

Description

Technical Field

[0001] This invention relates to the field of equipment fault diagnosis technology, specifically to a fault diagnosis method for wind turbines under time-varying operating conditions based on a spatiotemporal multi-scale graph convolutional neural network. Background Technology

[0002] With increasing global emphasis on sustainable development, new energy sources (such as wind and solar power) have become a crucial direction for future energy development. Wind power, as a green and environmentally friendly energy form, is gradually becoming an important component of the global energy structure transformation. However, during long-term operation, wind turbines may experience various malfunctions due to factors such as equipment aging and changes in the external environment, affecting their operating efficiency and the stability of power generation. Therefore, fault diagnosis of wind turbines is particularly important.

[0003] Supervisory Control and Data Acquisition (SCADA) systems are widely used in wind turbines. By collecting real-time data from various sensors on the wind turbine (such as temperature, vibration, and wind speed), they enable real-time monitoring of the equipment's operating status. SCADA systems provide a large amount of real-time data for fault diagnosis in wind farms. However, due to the complex spatiotemporal characteristics of wind turbine operating data, effectively extracting potential fault information from the data remains a major challenge for current fault diagnosis methods.

[0004] Currently, scholars have studied wind turbine fault diagnosis methods based on SCADA data. However, the time-varying operating conditions of wind turbines lead to strong instability in their fault characteristics, which vary across different time scales. Existing methods still do not adequately consider time-varying operating conditions and struggle to effectively capture and describe the correlation characteristics and dynamic coupling relationships of multiple sensors at different time scales. Therefore, many SCADA data-based wind turbine fault diagnosis methods struggle to deeply extract fault features and fail to achieve ideal diagnostic results. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a fault diagnosis method for time-varying operating conditions of wind turbine units based on spatiotemporal multi-scale graph convolutional neural networks, which addresses the shortcomings of the prior art.

[0006] The present invention adopts the following technical solution:

[0007] A fault diagnosis method for wind turbines under time-varying operating conditions based on spatiotemporal multi-scale graph convolutional neural networks includes the following steps:

[0008] S1. Collect samples of normal operation data and fault data from the historical SCADA monitoring data of the wind turbine unit;

[0009] S2. Perform data preprocessing on the normal operation data and fault data samples collected in step S1 to construct time features and adjacency matrices; step S2 includes the following steps:

[0010] S201. Label the normal data samples and faulty data samples;

[0011] S202. Construct a node feature matrix for the data after label processing, and calculate the distance matrix between nodes based on the Euclidean distance between node features;

[0012] S203. A nonlinear transformation is performed on the distance matrix using a Gaussian kernel function to construct an adjacency matrix. Self-loops are then added to the adjacency matrix, and threshold pruning and symmetric normalization are performed to generate forward and reverse adjacency matrices. The Gaussian kernel function calculation formula is as follows:

[0013]

[0014] in, yes Nodes at time steps and The similarity distance between nodes at time step t. yes and The Euclidean distance between them, where σ is the parameter of the Gaussian kernel function.

[0015] S204. Divide the data into time windows and generate time features. Use a fixed window length and step size to slide on the sample data to generate sample sequences. Each sample contains a fixed-length time series, its corresponding category label, and a time feature composed of the sliding window number and the time step position within the window.

[0016] S3. Construct a static predefined graph and a dynamic adaptive graph using the adjacency matrix and temporal features obtained in step S2, respectively, and fuse the graph convolution results using a gated fusion mechanism; step S3 includes the following steps:

[0017] S301. The two-dimensional temporal features and node spatial location features obtained in step S2 are mapped into spatiotemporal embedding vectors through a neural network. The nonlinear interaction of node attributes is modeled through a learnable matrix B to generate a dynamic adjacency matrix. The formula for generating the dynamic adjacency matrix is:

[0018]

[0019] in, It is the time step arrive The dynamic adjacency matrix between them and They are at the time step and spatiotemporal embedding, Used to remove weights smaller than a threshold, and using The function is normalized.

[0020] S302. Perform dynamic graph convolution on the dynamic adjacency matrix obtained in step S301 to extract dynamic adaptive spatial multi-scale features. The graph construction formula is as follows:

[0021]

[0022] in, It is the updated feature of layer l+1. These are the input features of the l-th layer. It is the time step arrive The adaptive adjacency matrix between them Represents time step arrive The adaptive adjacency matrix between them and These are learnable parameters.

[0023] S303. Using the adjacency matrix obtained in step S2, perform convolution feature extraction on the forward and reverse graphs respectively based on graph convolution operations to obtain static spatial multi-scale features. The graph construction formula is as follows:

[0024]

[0025] in, It is the updated feature of layer l+1. These are the input features of the l-th layer. and These are the forward adjacency matrix and the backward adjacency matrix, respectively. and These are learnable parameters.

[0026] S304. A gating fusion mechanism is used to fuse features extracted from the predefined graph and the adaptive graph. Gating weights are calculated by concatenating node features, and the sigmoid function is used to weight and fuse the features extracted from the static and dynamic graphs. The gating mechanism formula is as follows:

[0027]

[0028] in, The activation function is used to limit the value of the gate. It is a linear projection function that transforms feature vectors into one-dimensional vectors. and These are weight parameters.

[0029] S4. Using the graph convolution from step S3, perform four diffusing convolutions by expanding the causal structure to extract spatiotemporal features. The diffusing convolution formula is:

[0030]

[0031] in, It is the output feature. These are input features. It is a power series of the adjacency matrix. These are learnable parameters.

[0032] S5. The features obtained in step S4 at different scales are aggregated by calculating attention scores for node features at each scale, and the features at different time scales are weighted and aggregated. The fused spatiotemporal features are then used for classification and diagnosis.

[0033] By adopting the above technical solution, the present invention has the following beneficial effects:

[0034] This invention addresses the complex spatiotemporal characteristics of SCADA data by proposing a fault diagnosis method for wind turbines under time-varying operating conditions based on a spatiotemporal multi-scale graph convolutional neural network. The method models the spatial relationships between sensor nodes using a graph convolutional network (GCN) and captures multi-level dependencies of data at different time scales through dilated convolution techniques, thereby achieving accurate fault diagnosis. By combining a predefined graph structure with a dynamically adaptive graph structure, the relationship between nodes is dynamically adjusted according to data changes, enabling more flexible and efficient capture of spatiotemporal features. Furthermore, by incorporating multi-scale feature extraction, this method can simultaneously learn short-term anomalies and long-term trends in equipment operation at different time scales, significantly improving the accuracy and robustness of fault diagnosis. Attached Figure Description

[0035] Figure 1 This is a flowchart illustrating the implementation of the present invention;

[0036] Figure 2 This is a spatiotemporal expansion causal structure diagram of the present invention;

[0037] Figure 3 This is a schematic diagram of the spatiotemporal multi-scale graph convolutional neural network model of the present invention;

[0038] Figure 4 Diagnostic result diagram in an embodiment of the present invention. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] The SCADA data used in this embodiment of the invention spans from November 1, 2015 to January 1, 2016, with a wind turbine data sampling interval of 7 seconds. SCADA systems typically have hundreds of variables; this dataset, after processing, retains 26 continuous numerical variables. These variables cover multiple dimensions, including wind turbine operating parameters, environmental parameters, and status parameters.

[0041] See Figure 1 This document illustrates a flowchart of a method for diagnosing time-varying operating conditions of wind turbines using a spatiotemporal multi-scale graph neural network, as described in an embodiment of the present invention. This method utilizes a spatiotemporal multi-scale neural network model for wind turbine fault diagnosis. (See also...) Figure 3 This diagram illustrates a spatiotemporal multi-scale neural network model according to an embodiment of the present invention. The model combines a static predefined graph and a dynamic adaptive graph, and uses an dilated causal convolutional structure. The specific steps of this method are as follows:

[0042] S1. Collect normal operation data and fault data samples from the historical SCADA monitoring data of the wind turbine. In this embodiment, two types of samples are used: normal data and icing data. Variables include: wind speed, engine speed, yaw speed, etc.

[0043] S2. Perform data preprocessing on the normal operation data and fault data samples collected in step S1 to construct time features and adjacency matrices; step S2 includes the following steps:

[0044] S201. Label the normal data samples and faulty data samples;

[0045] S202. Construct a node feature matrix for the data after label processing, and calculate the distance matrix between nodes based on the Euclidean distance between node features;

[0046] S203. A nonlinear transformation is performed on the distance matrix using a Gaussian kernel function to construct an adjacency matrix. Self-loops are then added to the adjacency matrix, and threshold pruning and symmetric normalization are performed to generate forward and reverse adjacency matrices. The Gaussian kernel function calculation formula is as follows:

[0047]

[0048] in, yes Nodes at time steps and The similarity distance between nodes at time step t. yes and The Euclidean distance between them, where σ is the parameter of the Gaussian kernel function.

[0049] S204. Divide the data into time windows and generate time features. Use a fixed window length and step size to slide on the sample data to generate sample sequences. Each sample contains a fixed-length time series, its corresponding category label, and a time feature composed of the sliding window number and the time step position within the window.

[0050] S3. Construct a static predefined graph and a dynamic adaptive graph using the adjacency matrix and temporal features obtained in step S2, respectively, and fuse the graph convolution results using a gated fusion mechanism; step S3 includes the following steps:

[0051] S301. The two-dimensional temporal features and node spatial location features obtained in step S2 are mapped into spatiotemporal embedding vectors through a neural network. The nonlinear interaction of node attributes is modeled through a learnable matrix B to generate a dynamic adjacency matrix. The formula for generating the dynamic adjacency matrix is:

[0052]

[0053] in, It is the time step arrive The dynamic adjacency matrix between them and They are at the time step and spatiotemporal embedding, Used to remove weights smaller than a threshold, and using The function is normalized.

[0054] S302. Perform dynamic graph convolution on the dynamic adjacency matrix obtained in step S301 to extract dynamic adaptive spatial multi-scale features. The graph construction formula is as follows:

[0055]

[0056] in, It is the updated feature of layer l+1. These are the input features of the l-th layer. It is the time step arrive The adaptive adjacency matrix between them Represents time step arrive The adaptive adjacency matrix between them and These are learnable parameters.

[0057] S303. Using the adjacency matrix obtained in step S2, perform convolution feature extraction on the forward and reverse graphs respectively based on graph convolution operations to obtain static spatial multi-scale features. The graph construction formula is as follows:

[0058]

[0059] in, It is the updated feature of layer l+1. These are the input features of the l-th layer. and These are the forward adjacency matrix and the backward adjacency matrix, respectively. and These are learnable parameters.

[0060] S304. A gating fusion mechanism is used to fuse features extracted from the predefined graph and the adaptive graph. Gating weights are calculated by concatenating node features, and the sigmoid function is used to weight and fuse the features extracted from the static and dynamic graphs. The gating mechanism formula is as follows:

[0061]

[0062] in, The activation function is used to limit the value of the gate. It is a linear projection function that transforms feature vectors into one-dimensional vectors. and These are weight parameters.

[0063] S4. Using the graph convolution from step S3, perform four diffusing convolutions by expanding the causal structure to extract spatiotemporal features. The diffusing convolution formula is:

[0064]

[0065] in, It is the output feature. These are input features. It is a power series of the adjacency matrix. These are learnable parameters.

[0066] like Figure 2 As shown, by using an inflation factor k, selecting an input every k steps and performing a convolution operation on the selected input, it is possible to process longer time series with fewer layers.

[0067] S5. The features obtained in step S4 at different scales are fused using an attention mechanism, and the fused spatiotemporal features are used for classification and diagnosis; step S5 includes the following steps:

[0068] S501. Calculate the attention score for the node features at each scale, and perform weighted aggregation of features at different time scales to obtain the fused spatiotemporal features.

[0069] S502. Input the multi-scale spatiotemporal features into the softmax layer of the spatiotemporal multi-scale graph neural network model to complete the classification output.

[0070] This embodiment uses accuracy as the evaluation metric; higher accuracy indicates better diagnostic performance. This invention compares its diagnostic results with those of Convolutional Neural Networks (CNN), Convolutional Long Short-Term Memory Networks (CNN-LSTM), Convolutional Gated Unit Networks (CNN-GRU), and Spatiotemporal Graph Convolutional Neural Networks (STGCN), and the results are as follows: Figure 4 As shown, the diagnostic results of this invention are superior to those of other deep learning models.

[0071] The spatiotemporal multi-scale neural network-based wind turbine fault diagnosis method in this invention addresses the complex spatiotemporal characteristics of SCADA data by combining a static predefined graph with a dynamic adaptive graph. This approach mines the dynamic dependencies between different nodes at different time steps and utilizes an dilated causal convolution structure to extract features at different time scales. Compared to existing fault diagnosis models, this invention fully leverages the dynamic spatiotemporal correlations of SCADA data, enhances classification capabilities, and provides a novel technical solution for wind turbine fault diagnosis.

Claims

1. A method for fault diagnosis of wind turbine generators under time-varying operating conditions based on spatiotemporal multi-scale graph convolutional neural networks, characterized in that, Includes the following steps: S1. Collect samples of normal operation data and fault data from the historical SCADA monitoring data of the wind turbine unit; S2. Perform data preprocessing on the normal operation data and fault data samples collected in step S1 to construct time features and adjacency matrix; S3. Construct a static predefined graph and a dynamic adaptive graph using the adjacency matrix and time features obtained in step S2, and fuse the graph convolution results using a gated fusion mechanism. S4. Using the graph convolution from step S3, perform four diffusing convolutions by expanding the causal structure to extract spatiotemporal features. The diffusing convolution formula is: in, It is the output feature. These are input features. It is a power series of the adjacency matrix. These are learnable parameters; S5. The features obtained in step S4 at different scales are aggregated by calculating attention scores for node features at each scale, and the features at different time scales are weighted and aggregated. The fused spatiotemporal features are then used for classification and diagnosis.

2. The method for fault diagnosis of wind turbines under time-varying operating conditions based on spatiotemporal multi-scale graph convolutional neural networks according to claim 1, characterized in that, Step S2 includes the following steps: S201. Label the normal data samples and faulty data samples; S202. Construct a node feature matrix for the data after label processing, and calculate the distance matrix between nodes based on the Euclidean distance between node features; S203. A nonlinear transformation is performed on the distance matrix using a Gaussian kernel function to construct an adjacency matrix. Self-loops are then added to the adjacency matrix, and threshold pruning and symmetric normalization are performed to generate forward and reverse adjacency matrices. The Gaussian kernel function calculation formula is as follows: in, yes Nodes at time steps and The similarity distance between nodes at time step t. yes and The Euclidean distance between them, where σ is the parameter of the Gaussian kernel function; S204. Divide the data into time windows and generate time features. Use a fixed window length and step size to slide on the sample data to generate sample sequences. Each sample contains a fixed-length time series, its corresponding category label, and a time feature composed of the sliding window number and the time step position within the window.

3. The method for fault diagnosis of wind turbines under time-varying operating conditions based on spatiotemporal multi-scale graph convolutional neural networks according to claim 2, characterized in that, The temporal features of the SCADA data in step S204 include "within the sliding window" and "between the sliding window". After one-hot encoding, the data is embedded in a trainable fully connected network for dimensionality reduction.

4. The method for fault diagnosis of wind turbines under time-varying operating conditions based on spatiotemporal multi-scale graph convolutional neural networks according to claim 1, characterized in that, Step S3 includes the following steps: S301. The two-dimensional temporal features and node spatial location features obtained in step S2 are mapped into spatiotemporal embedding vectors through a neural network. The nonlinear interaction of node attributes is modeled through a learnable matrix B to generate a dynamic adjacency matrix. The formula for generating the dynamic adjacency matrix is: in, It is the time step arrive The dynamic adjacency matrix between them and They are at the time step and spatiotemporal embedding, Used to remove weights smaller than a threshold, and using Normalize the function; S302. Perform dynamic graph convolution on the dynamic adjacency matrix obtained in step S301 to extract dynamic adaptive spatial multi-scale features. The graph construction formula is as follows: in, It is the update feature of the (l+1)th layer. These are the input features of the l-th layer. It is the time step arrive The adaptive adjacency matrix between them Represents time step arrive The adaptive adjacency matrix between them and These are learnable parameters; S303. Using the adjacency matrix obtained in step S2, perform convolution feature extraction on the forward and reverse graphs respectively based on graph convolution operations to obtain static spatial multi-scale features. The graph construction formula is as follows: in, It is the update feature of the (l+1)th layer. These are the input features of the l-th layer. and These are the forward adjacency matrix and the backward adjacency matrix, respectively. and These are learnable parameters; S304. A gating fusion mechanism is used to fuse features extracted from the predefined graph and the adaptive graph. Gating weights are calculated by concatenating node features, and the sigmoid function is used to weight and fuse the features extracted from the static and dynamic graphs. The gating mechanism formula is as follows: in, The activation function is used to limit the value of the gate. It is a linear projection function that transforms feature vectors into one-dimensional vectors. and These are weight parameters.

Images