Wind turbine generator fault diagnosis method based on space-time diagram convolutional neural network

Abstract

The invention discloses a wind turbine generator fault diagnosis method based on a space-time diagram convolutional neural network, and belongs to the technical field of equipment fault diagnosis, and the method comprises the following steps: data collection, data preprocessing, data feature extraction and fusion, and fault diagnosis result output. Aiming at complex space-time characteristics of SCADA data, a space-time diagram convolutional neural network combining a one-dimensional convolutional neural network and a diagram convolutional neural network is provided, the one-dimensional convolutional neural network aims at time sequence characteristics, and the diagram convolutional neural network aims at space characteristics. The time-space features of the SCADA data are learned through the time-space diagram convolutional neural network model combined with two networks, and the fault diagnosis result is given, so that the fault diagnosis performance is improved, and the problem of low prediction accuracy of the existing wind turbine generator fault diagnosis model is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment fault diagnosis, and in particular to a wind turbine fault diagnosis method based on spatiotemporal graph convolutional neural network. Background Art

[0002] As the global demand for renewable energy continues to grow, wind power generation has been widely used around the world as a clean and sustainable form of energy. Wind turbines are the core equipment of wind power generation systems, and their operating status and performance directly affect the power generation efficiency and economic benefits of wind farms. However, the operating environment of wind turbines is complex and is affected by multiple factors such as wind speed, temperature, and humidity, making them prone to various faults. Traditional wind turbine fault diagnosis methods mainly rely on regular maintenance and manual inspections, which have problems such as long detection cycles, low efficiency, and high costs. Therefore, how to efficiently monitor and diagnose wind turbine faults and ensure their stable operation has become a technical problem that urgently needs to be solved.

[0003] In modern wind farms, supervisory control and data acquisition (SCADA) systems are widely deployed to collect various operating data of wind turbines in real time. These data include wind speed, rotation speed, power output, temperature, vibration and other parameters, forming a multi-dimensional, time-series, and huge data set. These SCADA data not only record the operating status of wind turbines, but also contain rich fault characteristic information. Therefore, how to efficiently process and analyze these SCADA data has become the key to improving the fault diagnosis capabilities of wind turbines.

[0004] With the development of sensor technology and the Internet of Things, spatiotemporal data collected by SCADA systems can be used to establish data-driven fault diagnosis models. However, traditional data processing methods often only focus on the time dimension or the space dimension when processing SCADA data, which makes it difficult to fully capture the complex relationships in the data, limiting the accuracy of fault diagnosis and prediction. Summary of the invention

[0005] The present invention provides a wind turbine fault diagnosis method based on spatiotemporal graph convolutional neural network to solve the problem that the prediction accuracy of the existing wind turbine fault diagnosis model is low.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

[0007] A wind turbine fault diagnosis method based on spatiotemporal graph convolutional neural network includes the following steps:

[0008] S1. Data collection: collecting normal operation data samples and fault data samples from historical SCADA data.

[0009] S2, data preprocessing, performing data preprocessing on the normal operation data samples and fault data samples collected in step S1 to obtain a time window and an adjacency matrix; wherein step S2 includes the following steps:

[0010] S201, labeling normal operation data samples and fault data samples;

[0011] S202, performing feature selection on the labeled data to obtain a data set;

[0012] S203, dividing the data set into time windows to obtain a time window with the minimum unit of model input;

[0013] S204, according to the data set, the similarity distance between each sensor data type is calculated by the Gaussian kernel function, and an adjacency matrix is ​​constructed; the similarity distance between each sensor data type is calculated using the complete data set, thereby constructing an adjacency matrix; the Gaussian kernel function calculation formula is:

[0014]

[0015] in, is x i and x j The similarity distance between two features, x i and x j are feature vectors, representing different sensor data, σ are the parameters of the Gaussian kernel function;

[0016] The division of the time window is coordinated with the size of the convolution kernel of the one-dimensional convolution in the one-dimensional convolutional neural network, and an appropriate time window length is selected so that the model performs a certain number of time dimension convolutions within the time window;

[0017] S3, data feature extraction and fusion, input the adjacency matrix and time window obtained in step S2 into the spatiotemporal graph convolutional neural network model, extract temporal features and spatial multi-scale features from the adjacency matrix through a one-dimensional convolutional neural network and a graph convolutional neural network, respectively, fuse the extracted features to obtain spatiotemporal fusion features; wherein step S3 includes the following steps:

[0018] S301, performing one-dimensional convolution on the data of the adjacency matrix in the time dimension within each time window to extract time series features. The calculation formula of the one-dimensional convolution is:

[0019]

[0020] Among them, y(t) represents the value of the output signal at time t, k represents the convolution kernel of the one-dimensional convolution, and x(a) represents the input signal at time aThe value at time ta, k(ta) represents the value of the convolution kernel at time ta;

[0021] S302: Perform graph convolution on the spatial feature dimension of the data extracted from the time series features within the time window to extract the spatial features. The calculation formula for graph convolution is:

[0022]

[0023] Among them, H (l) is the input feature of the lth layer, H (l+1) is the feature of the l+1th layer is the normalized adjacency matrix, is the degree matrix, the purpose is to give the adjacency matrix Add self-loop, W (l) The weight matrix of the lth layer, σ is the activation function;

[0024] S303, repeating step S301, performing temporal feature extraction again on the data after spatial feature extraction, and steps S301-S303 constitute a spatiotemporal feature extraction process;

[0025] S304, repeating steps S301-S303, and performing the spatiotemporal feature extraction process again;

[0026] S305, performing one-dimensional convolution on the time axis dimension again on the data that has undergone two spatiotemporal feature extraction processes;

[0027] S306, the data features processed in step S305 are sequentially subjected to global average pooling and a fully connected layer for dimensionality reduction to obtain the spatiotemporal fusion features.

[0028] S4: Output the fault diagnosis result. Input the spatiotemporal fusion features obtained in step S3 into the softmax layer in the spatiotemporal graph convolutional network model to obtain the final wind turbine fault diagnosis result. The Softmax function formula is:

[0029]

[0030] in, z are input features, is the probability of the i-th class, and K is the total number of classes.

[0031] Due to the adoption of the above technical solution, the present invention has the following beneficial effects:

[0032] The present invention aims at the complex spatiotemporal characteristics of SCADA data, and proposes a spatiotemporal graph convolutional neural network that combines a one-dimensional convolutional neural network with a graph convolutional neural network, wherein the one-dimensional convolutional neural network is for the time series characteristics, and the graph convolutional neural network is for the space characteristics. The spatiotemporal characteristics of SCADA data are learned by the spatiotemporal graph convolutional neural network model that combines the two networks, and a fault diagnosis result is given, thereby improving the performance of fault diagnosis. The present invention can fully extract the complex spatiotemporal characteristics in SCADA data and greatly improve the accuracy of fault diagnosis. Compared with existing fault diagnosis models (for example: long short-term memory network (LSTM), convolutional neural network (CNN), convolutional neural network-gated recurrent unit (CNN-GRU), multi-scale spatiotemporal attention mechanism (MT-STAN), etc.), the present invention has better performance in classification performance indicators, can more accurately capture the changes in the state of wind turbine generator sets, and thus improve the effect of fault diagnosis, and provides a new technical method for the field of wind turbine fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a flow chart for realizing the present invention;

[0034] Figure 2 It is a schematic diagram of the input time window of the present invention;

[0035] Figure 3 A one-dimensional convolutional neural network (1D-CNN) diagram of the present invention;

[0036] Figure 4 is a graph convolutional neural network (GCN) graph of the present invention;

[0037] Figure 5 Schematic diagram of the spatiotemporal graph convolutional neural network model of the present invention;

[0038] Figure 6 This is a comparison chart of the accuracy of the present invention and other deep learning fault diagnosis models. DETAILED DESCRIPTION

[0039] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0040] Example

[0041] like Figure 1 As shown, a wind turbine fault diagnosis method based on spatiotemporal graph convolutional neural network includes steps S1-S4.

[0042] S1, data collection, collecting normal operation data samples and fault data samples in historical SCADA data. In this embodiment, two samples of normal data and icing data are used, and the variables include: wind speed, engine speed, yaw position, yaw speed, etc.

[0043] S2, data preprocessing, performing data preprocessing on the normal operation data samples and fault data samples collected in step S1 to obtain a time window and an adjacency matrix; wherein step S2 includes the following steps:

[0044] S201: label normal operation data samples and fault data samples.

[0045] S202: Perform feature selection on the labeled data to obtain a data set. Not all sensor data may be used.

[0046] S203: Divide the data set into time windows to obtain a time window of the minimum unit of model input.

[0047] S204, according to the data set, the similarity distance between each sensor data type is calculated by the Gaussian kernel function, and an adjacency matrix is ​​constructed. This process uses the complete data set to calculate the similarity distance between each sensor data type, thereby constructing an adjacency matrix. The Gaussian kernel function calculation formula is:

[0048]

[0049] in, is x i and x j The similarity distance between two features, x i and x j are feature vectors, representing different sensor data, σ are the parameters of the Gaussian kernel function.

[0050] In addition, the division of the time window is coordinated with the convolution kernel size of the one-dimensional convolution in the one-dimensional convolutional neural network, and an appropriate time window length is selected so that the model performs a certain number of time dimension convolutions (for example, 5 times) within the time window.

[0051] S3, data feature extraction and fusion, input the adjacency matrix and time window obtained in step S2 into the spatiotemporal graph convolutional neural network model, extract temporal features and spatial multi-scale features from the adjacency matrix through a one-dimensional convolutional neural network and a graph convolutional neural network, respectively, fuse the extracted features to obtain spatiotemporal fusion features; wherein step S3 includes the following steps:

[0052] S301, performing one-dimensional convolution on the data of the adjacency matrix in the time dimension within each time window to extract time series features. The calculation formula of the one-dimensional convolution is:

[0053]

[0054] Among them, y(t) represents the value of the output signal at time t, k represents the convolution kernel of the one-dimensional convolution, and x(a) represents the input signal at time a The value at time ta, k(ta) represents the value of the convolution kernel at time ta;

[0055] S302: Perform graph convolution on the spatial feature dimension of the data extracted from the time series features within the time window to extract the spatial features. The calculation formula for graph convolution is:

[0056]

[0057] Among them, H (l) is the input feature of the lth layer, H (l+1) is the feature of the l+1th layer is the normalized adjacency matrix, is the degree matrix, the purpose is to give the adjacency matrix Add self-loop, W (l) The weight matrix of the lth layer, σ is the activation function;

[0058] S303, repeating step S301, performing temporal feature extraction again on the data after spatial feature extraction, and steps S301-S303 constitute a spatiotemporal feature extraction process;

[0059] S304, repeating steps S301-S303, and performing the spatiotemporal feature extraction process again;

[0060] S305, performing one-dimensional convolution on the time axis dimension again on the data that has undergone two spatiotemporal feature extraction processes;

[0061] S306, the data features processed in step S305 are sequentially subjected to global average pooling and fully connected layers for dimension reduction to obtain the spatiotemporal fusion features, retaining important information in preparation for classification.

[0062] The main function of step S305 is to extract the features of the data, while the main function of step S306 is to reduce the dimensionality of the data so as to make predictions in the next step.

[0063] like Figure 2As shown in the figure, each node graph represents the association between all sensor features at each time point. The data contains both the temporal characteristics of the time dimension and the spatial characteristics of each node graph. The feature extraction process is to perform one-dimensional convolutional neural network convolution in the time axis dimension and graph convolution on all features on each node graph (each time point) to extract spatial characteristics.

[0064] like Figure 3-5 As shown in the figure, after performing a time dimension convolution, the graph convolution neural network is used to perform graph convolution on the node graph at each time point, and the feature vector is multiplied by the adjacency matrix for feature aggregation, and then the spatial features are fully extracted through feature learning to obtain a new feature vector. The process of alternating two time dimension convolutions and one graph convolution is called a spatiotemporal feature extraction process. After two spatiotemporal feature extraction processes, the spatiotemporal features are fully extracted and then another time dimension convolution is performed.

[0065] S4: Output the fault diagnosis result. Input the spatiotemporal fusion features obtained in step S3 into the softmax layer in the spatiotemporal graph convolutional network model to obtain the final wind turbine fault diagnosis result. The Softmax function formula is:

[0066]

[0067] in, z are input features, is the probability of the i-th class, and K is the total number of classes.

[0068] This step inputs the reduced features into the softmax layer to complete the classification decision. The model selects the highest probability category as the final prediction result.

[0069] This example uses accuracy as an evaluation index. The closer the evaluation index is to 100%, the better the diagnostic effect of the surface method. In order to illustrate the superiority of the present invention, this example uses a long short-term memory network (LSTM), a convolutional neural network (CNN), a CNN-GRU (convolutional neural network-gated recurrent unit), and a spatiotemporal attention mechanism (STAM) for comparison. Figure 6 The comparison results of the models on the classification performance evaluation indicators are shown. It can be seen that the diagnosis results of the present invention are better than those of other deep learning models.

[0070] The present invention aims at the complex spatiotemporal characteristics of SCADA data, and proposes a spatiotemporal graph convolutional neural network that combines a one-dimensional convolutional neural network with a graph convolutional neural network, wherein the one-dimensional convolutional neural network is for the time series characteristics, and the graph convolutional neural network is for the spatial characteristics. The spatiotemporal characteristics of SCADA data are learned through the spatiotemporal graph convolutional neural network model that combines the two networks, and a fault diagnosis result is given, thereby improving the performance of fault diagnosis. The present invention can fully extract the complex spatiotemporal characteristics in SCADA data and greatly improve the accuracy of fault diagnosis. Compared with the existing fault diagnosis model, the present invention has better performance in classification performance indicators, can more accurately capture the changes in the state of wind turbines, and thus improve the effect of fault diagnosis, providing a new technical method for the field of wind turbine fault diagnosis.

[0071] The above description is a detailed description of the preferred feasible embodiments of the present invention, but the embodiments are not intended to limit the scope of the patent application of the present invention. All equivalent changes or modified changes completed under the technical spirit suggested by the present invention should fall within the patent scope covered by the present invention.

Claims

1. A wind turbine fault diagnosis method based on spatiotemporal graph convolutional neural network, characterized in that: The following steps are involved: S1. Data collection: collecting normal operation data samples and fault data samples from historical SCADA data; S2, data preprocessing, performing data preprocessing on the normal operation data samples and fault data samples collected in step S1 to obtain a time window and an adjacency matrix; S3, data feature extraction and fusion, input the adjacency matrix and time window obtained in step S2 into the spatiotemporal graph convolutional neural network model, perform temporal feature extraction and spatial multi-scale feature extraction on the adjacency matrix through a one-dimensional convolutional neural network and a graph convolutional neural network respectively, fuse the extracted features to obtain spatiotemporal fusion features; S4: Output the fault diagnosis result. Input the spatiotemporal fusion features obtained in step S3 into the softmax layer in the spatiotemporal graph convolutional network model to obtain the final wind turbine fault diagnosis result. The Softmax function formula is: in, z are input features, is the probability of the i-th class, and K is the total number of classes.

2. A wind turbine fault diagnosis method based on spatiotemporal graph convolutional neural network according to claim 1, characterized in that: Step S2 includes the following steps: S201, labeling normal operation data samples and fault data samples; S202, performing feature selection on the labeled data to obtain a data set; S203, dividing the data set into time windows to obtain a time window with the minimum unit of model input; S204. According to the data set, the similarity distance between each sensor data type is calculated by using the Gaussian kernel function, and an adjacency matrix is ​​constructed. The calculation formula of the Gaussian kernel function is: in, is x i and x j The similarity distance between two features, x i and x j are feature vectors, representing different sensor data, σ are the parameters of the Gaussian kernel function.

3. A wind turbine fault diagnosis method based on spatiotemporal graph convolutional neural network according to claim 2, characterized in that: In step S203, the division of the time window is coordinated with the convolution kernel size of the one-dimensional convolution in the one-dimensional convolutional neural network, and an appropriate time window length is selected so that the model performs a certain number of time dimension convolutions within the time window.

4. A wind turbine fault diagnosis method based on spatiotemporal graph convolutional neural network according to claim 3, characterized in that: In step S204, the complete data set is used to calculate the similarity distance between each sensor data type, thereby constructing an adjacency matrix.

5. The wind turbine fault diagnosis method based on spatiotemporal graph convolutional neural network according to claim 1 is characterized in that: Step S3 includes the following steps: S301, performing one-dimensional convolution on the data of the adjacency matrix in the time dimension within each time window to extract time series features. The calculation formula of the one-dimensional convolution is: Among them, y(t) represents the value of the output signal at time t, k represents the convolution kernel of the one-dimensional convolution, and x(a) represents the input signal at time a The value at time ta, k(ta) represents the value of the convolution kernel at time ta; S302: Perform graph convolution on the spatial feature dimension of the data extracted from the time series features within the time window to extract the spatial features. The calculation formula for graph convolution is: Among them, H (l) is the input feature of the lth layer, H (l+1) is the feature of the l+1th layer is the normalized adjacency matrix, is the degree matrix, the purpose is to give the adjacency matrix Add self-loop, W (l) The weight matrix of the lth layer, σ is the activation function; S303, repeating step S301, performing temporal feature extraction again on the data after spatial feature extraction, and steps S301-S303 constitute a spatiotemporal feature extraction process; S304, repeating steps S301-S303, and performing the spatiotemporal feature extraction process again; S305, performing one-dimensional convolution on the time axis dimension again on the data that has undergone two spatiotemporal feature extraction processes; S306, the data features processed in step S305 are sequentially subjected to global average pooling and a fully connected layer for dimensionality reduction to obtain the spatiotemporal fusion features.

Images