A fan gearbox working condition anomaly identification method based on node-causal inference network

Abstract

The application discloses a fan gearbox working condition anomaly identification method based on a node-causal reasoning network and belongs to the technical field of working condition anomaly identification. The method comprises the following steps: S1: obtaining an original time sequence of fan gearbox fault data, segmenting the original time sequence to obtain a time sequence data slice, and performing fault labeling on the time sequence data slice; S2: decomposing the time sequence data slice to obtain a data block component; S3: generating a fault feature coupling graph based on the data block component; and S4: inputting the fault feature coupling graph and the fault labeling into a support vector machine to obtain a working condition anomaly identification result. The application is superior to a traditional method in the performance on a fan gearbox fault data set, can well solve the working condition anomaly identification problem of the fan gearbox, and has good stability in the performance on a standard time sequence classification data set.

Description

Technical Field

[0001] This invention belongs to the field of operating condition anomaly identification technology, specifically relating to a method for identifying operating condition anomalies in wind turbine gearboxes based on node-causal reasoning networks. Background Technology

[0002] With the significant increase in the scale of wind turbine installations and grid-connected capacity, research on efficient operation and maintenance and anomaly identification is of great significance for the safe and stable operation of wind turbines. The gearbox, the most critical component in the wind turbine transmission system, frequently experiences faults such as cracking and spalling due to its harsh working environment; these faults account for approximately 59% of all turbine failures. Therefore, timely and accurate diagnosis of wind turbine gearbox faults is crucial for improving the efficiency and reliability of wind turbines and reducing operation and maintenance costs.

[0003] The randomness of wind energy causes wind turbine gearboxes to operate under non-stationary conditions year-round. Under alternating load excitation, the unit fault signals exhibit significant nonlinear time-varying characteristics. At the same time, considering the convolutional coupling effect of complex environmental noise in the transmission path, traditional wind turbine gearbox vibration signal processing methods pose a great challenge. Therefore, how to model this problem and effectively identify abnormal operating conditions has become a difficult problem.

[0004] Vibration signals are a rich source of fault information and have been widely used in the study of abnormal operating conditions in wind turbine gearboxes. Currently, intelligent operating condition anomaly identification methods based on vibration signals have become a hot topic in the research of non-stationary signals in wind turbine gearboxes. Generally, intelligent operating condition anomaly identification methods mainly consist of two steps: feature extraction and fault classification. However, existing intelligent operating condition anomaly identification methods have limited learning capabilities for complex signals and cannot effectively learn the potential features of fault signals. Furthermore, fault identification under varying operating conditions is even more difficult. Research on time series classification has made significant progress in recent years, resulting in a series of valuable methods in many application areas. However, traditional time series classification algorithms have limitations when processing noisy time series data. Summary of the Invention

[0005] The purpose of this invention is to solve the time series classification problem of wind turbine gearboxes, overcome the limitations of traditional methods, and propose a time series feature extraction method based on temporal phase-energy decomposition and node-causal inference network. Then, support vector machine is used to classify the extracted feature matrix.

[0006] To achieve the above objectives, the present invention provides the following solution: a method for identifying abnormal operating conditions of wind turbine gearboxes based on node-causal reasoning networks, comprising the following steps: S1: Obtain the original time series of wind turbine gearbox fault data, segment the original time series to obtain time series data slices, and label the time series data slices with faults. S2: Decompose the time series data slices to obtain data block components; S3: Generate a fault feature coupling map based on the data block components; S4: Input the fault feature coupling map and fault labeling into the support vector machine to obtain the abnormal operating condition identification result.

[0007] More preferably, the time series data slice is ,in, This represents the number of slices in a one-dimensional time series data. Indicates the length of each slice of time series data; Labels representing slices of time series data; R Represents the set of real numbers; The matrix form is as follows: ; ; In the formula, Indicates the first A slice of time series data.

[0008] More preferably, the data block component includes: ; in, ; In the formula, p Indicates the number of data block components; Represents a time series; n Indicates the length of a one-dimensional time series; L Indicates the length of the lag vector sequence; K = N-L +1; Indicates the length of each slice of time series data; x i,j The first time series i Okay, number j The elements of the column.

[0009] More preferably, in S3, the problem of learning the fault feature coupling map is transformed into solving a least squares problem, and the objective function is constructed as follows: ; In the formula, W j The first part represents the fault feature coupling spectrum.j List; β >0 indicates the regularization parameter; Z Indicates the independent variable; Y j This represents the dependent variable.

[0010] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention models the problem of identifying abnormal operating conditions in wind turbine gearboxes as a time-series classification task. The original fault data is segmented into numerous time-series data blocks, and expert knowledge is used to label the fault types of these blocks. A univariate time-series classification method for identifying abnormal operating conditions in wind turbine gearboxes is then proposed. Combining the dimensionality increase and noise reduction capabilities of temporal phase-energy decomposition with the feature extraction and fitting capabilities of node-causal inference networks, a convex optimization method is used to learn the node-causal inference network, accelerating its training time. Finally, a support vector machine is used to classify the feature matrix. The effectiveness of the proposed classification method is verified through experiments on a fault dataset of wind turbine gearboxes in a real wind farm and a standard time-series classification dataset. The proposed method outperforms traditional methods on the wind turbine gearbox fault dataset, effectively solving the problem of identifying abnormal operating conditions in wind turbine gearboxes, and exhibits good stability on the standard time-series classification dataset. Attached Figure Description

[0011] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a schematic diagram of the process for identifying abnormal operating conditions of wind turbine gearboxes based on node-causal reasoning networks, provided in an embodiment of the present invention. Detailed Implementation

[0013] 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.

[0014] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0015] Example 1: like Figure 1 As shown, this embodiment provides a method for identifying abnormal operating conditions of wind turbine gearboxes based on node-causal reasoning networks, including the following steps: S1: Obtain the original time series of wind turbine gearbox fault data, segment the original time series to obtain time series data slices, and label the time series data slices with faults.

[0016] Segmenting the original time series facilitates data analysis for each time period, allowing for the acquisition of more fault feature coupling maps. Labeling the faults facilitates subsequent classification of different data types, resulting in time series data slices. ,in, This represents the number of slices in a one-dimensional time series data. Indicates the length of each slice of time series data; Labels representing slices of time series data; R It represents the set of real numbers. The matrix form is as follows: ; ; In the formula, Indicates the first A slice of time series data.

[0017] To avoid the impact of scale differences between different data on model training performance, the time series data slices are normalized to unify the data to the same range.

[0018] ; In the formula, It is sample data. It is the minimum value of the sample. It is the maximum value of the sample.

[0019] S2: The time series data slices are decomposed using the temporal phase-energy decomposition method to obtain data block components.

[0020] The length is N Time series data slices Convert to length of L The lag vector sequence, 1< L < N ,form K A length of L The vector, where, K = N - L +1. This leads to the trajectory matrix: .

[0021] Each column of the trajectory matrix consists of a length of The sliding window is generated from slices of time series data. It is the only parameter, and its selection directly affects the size and effectiveness of the temporal phase-energy decomposition, and even more directly affects the subsequent classification results. Its value is determined based on the specific time series.

[0022] The trajectory matrix is ​​decomposed into a series of biorthogonal elementary matrices of rank 1 using singular value decomposition: ; In the formula, U The left singular vector matrix representing the trajectory matrix; V The right singular vector matrix representing the trajectory matrix; The rank of the trajectory matrix is ​​( ). e i Singular values ​​are represented in descending order; u i and v i These represent the left and right singular vectors of the trajectory matrix.

[0023] Therefore, the trajectory matrix is ​​decomposed into: .

[0024] Will Divided into p Given three distinct subsets, sum the matrix over each subset as follows: .

[0025] along The purpose of averaging the cross diagonals is to find the closest... Least squared number of bits in the trajectory matrix and time series Time series The elements are derived from the following formula: ; In the formula, n Indicates the length of a one-dimensional time series; x i,j The first of the inter-sequences i Okay, number j Column elements.

[0026] Finally, the entire process described above slices the time series data. H Expand to meet the requirements Data block components: .

[0027] Gearbox fault data is one-dimensional time series data. Through the above processing, within the window length... Under certain conditions, the denoised data block components (including main trend components and periodic components, etc.) can be obtained and used as input to the node-causal inference network to learn the fault feature coupling map of the fault data.

[0028] S3: Generate a fault feature coupling map based on the data block components.

[0029] From the above p Select from the components N c Each data block component serves as a state inference node in the node-causal inference network, yielding a time series for fault feature coupling graph learning. : ; In the formula, u ij express i The state reasoning node at the _th ... j Time (or the) j The state value of the next iteration.

[0030] No.( t +1) The state reasoning node causal value of the next iteration is coupled to the fault feature graph. W and the t The state value is determined in the next iteration, and the update of the state value is expressed by the following formula: C i ( t +1)= f ( C ( t ) W j ); In the formula, This represents the state values ​​of all state inference nodes in the node-causal inference network at time t. C i ( t +1) indicates the node-causal reasoning network's first... j Each state reasoning node is in t The state value at time +1; Representing the coupling spectrum of fault characteristics W The j List; f This represents the activation function used to limit the state value to the interval [0,1]. In this embodiment, the Sigmoid function is used.

[0031] Performing an inverse transform on the Sigmoid function yields: ; In the formula, Lambda This represents the steepness parameter of the Sigmoid activation function, used to adjust the slope of the Sigmoid function; y This represents the state value after activation by the Sigmoid function.

[0032] Performing an inverse transformation on the state values ​​yields: .

[0033] make ,in, ; ,Will Z and Y j Substituting into the above equation, we get: Y j = ZW j .

[0034] Z and Y j These can be considered as independent and dependent variables, respectively. Therefore, the problem of learning the coupled fault feature map is transformed into solving a least squares problem. To solve... W j Construct the following objective function: ; In the formula, β >0 indicates the regularization parameter.

[0035] By minimizing the least squares error between the actual and predicted values ​​using the objective function, we can obtain... W j A reasonable approximate solution. Constraint terms. This ensures that each value in the fault feature coupling map is within the interval [-1, 1]. Clearly, the objective function described above is a standard convex optimization problem with linear constraints. Several convex optimization techniques exist for solving this problem, such as the commonly used interior-point method and the primal-dual method. This process is repeated for all nodes to generate the fault feature coupling map for each time block component. All fault feature coupling maps are then merged into... M ∈ R （Nc×（Nc×S）） .

[0036] S4: Input the fault feature coupling map and fault labeling into the support vector machine for abnormal operating condition identification.

[0037] This implementation uses a classic Support Vector Machine (SVM) classifier to perform multi-class classification of the fault feature matrix. The multi-class method trains a classifier between every two classes, and the final attributes of the test data are determined by voting. Specifically, the libsvm toolbox in MATLAB is used for multi-class classification. After feature extraction from the time series data slices, the generated data is represented as follows: M ={ W i , i =1,2,..., S , Y ∈ R S}

[0038] Will S The group fault feature coupling map and its corresponding labels are set as input to the SVM. The SVM is used to classify the data to obtain the results of abnormal operating conditions.

[0039] Example 2: The method proposed in this invention was applied to a real-world wind turbine gearbox failure dataset and a laboratory wind turbine gearbox failure dataset to demonstrate its superior performance. Simultaneously, the effectiveness of the invention was verified using 10 standard publicly available time-series classification datasets. All datasets were divided into training and testing sets. No data processing was performed before training the model. Detailed information on the wind turbine gearbox failure dataset and the standard publicly available datasets is shown in Tables 1 and 2.

[0040] Table 1 Table 2 The actual gearbox data from a wind farm in Jilin Province comes from wind turbine units in a real wind farm. This data contains four types of faults. The fault time series is divided into 904 parts (226 faults per part), with each part containing 16,384 sampling points. 640 parts are used as the training set, and 264 parts are used as the test set. The laboratory gearbox data comes from the Qianpeng QPZZ-II anomaly identification system. This data contains three types of faults and normal states. The fault time series is divided into 400 parts, with each part containing 500 sampling points. 280 parts are used as the training set, and 120 parts are used as the test set.

[0041] In this experiment, temporal phase-energy decomposition was first used to decompose and increase the dimensionality of each time series block. Then, convex optimization was used to learn the fault feature coupling map of each time series block. Finally, SVM was used to classify the fault feature coupling map of each time series block. The activation function of the node-causal inference network is the Sigmoid function, and all parameters are set to 5, so that the value of each node is controlled within [...]. Between 1, 1]. The values ​​of the wind turbine gearbox failure time series are mapped to [ ] using the following formula. Between 1 and 1].

[0042] First, the performance of the proposed method was compared with other basic classification methods on a wind turbine gearbox fault dataset. In this embodiment, the basic classification methods include: 1NN, decision tree, random forest, and SVM, as shown in Tables 3 and 4. Table 3 shows the comparison results of the proposed method with other methods on actual wind turbine gearbox fault data from a wind farm; Table 4 shows the comparison results of the proposed method with other methods on laboratory wind turbine gearbox fault data. The classification accuracy of the proposed method has a significant advantage over other methods, mainly attributed to the noise filtering effect of temporal phase-energy decomposition, the ability to extract useful features, and the fitting ability of the node-causal inference network. The specific parameters of the proposed method are set as follows: actual wind turbine gearbox fault dataset. L =5, N c =5, and achieved the highest accuracy of 93.56% under this parameter; Laboratory fan gearbox fault dataset L =4, N c =4, and the highest accuracy rate of 91.67% is achieved under this parameter.

[0043] Table 3 Table 4 The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for identifying abnormal operating conditions of wind turbine gearboxes based on node-causal reasoning networks, characterized in that, Includes the following steps: S1: Obtain the original time series of wind turbine gearbox fault data, segment the original time series to obtain time series data slices, and label the time series data slices with faults. S2: Decompose the time series data slices to obtain data block components; S3: Generate a fault feature coupling map based on the data block components; S4: Input the fault feature coupling map and fault labeling into the support vector machine to obtain the abnormal operating condition identification result.

2. The method for identifying abnormal operating conditions of wind turbine gearboxes based on node-causal reasoning networks according to claim 1, characterized in that, The time series data slice is ,in, This represents the number of slices in a one-dimensional time series data. Indicates the length of each slice of time series data; Labels representing slices of time series data; R Represents the set of real numbers; The matrix form is as follows: ； ； In the formula, Indicates the first A slice of time series data.

3. The method for identifying abnormal operating conditions of wind turbine gearboxes based on node-causal reasoning networks according to claim 1, characterized in that, The data block components include: ； in, ； In the formula, p Indicates the number of data block components; Represents a time series; n Indicates the length of a one-dimensional time series; L Indicates the length of the lag vector sequence; K = N - L +1; Indicates the length of each slice of time series data; x i,j The first time series i Okay, number j The elements of the column.

4. The method for identifying abnormal operating conditions of wind turbine gearboxes based on node-causal reasoning networks according to claim 1, characterized in that, In S3, the problem of learning the fault feature coupling map is transformed into solving a least squares problem, and the objective function is constructed as follows: ； In the formula, W j The first part represents the fault feature coupling map. j List; β >0 indicates the regularization parameter; Z Indicates the independent variable; Y j This represents the dependent variable.

Images