Fan blade icing detection method and system based on spatiotemporal attention twin network

Abstract

The application discloses a fan blade icing detection method and system based on a space-time attention twin network, belongs to the field of fan equipment fault diagnosis, and performs pretreatment on multivariate sensor data collected from a wind farm SCADA system, divides a training set, a verification set and a test set, constructs an attribute graph according to similarity information between target sensor variable data, and uses the attribute graph as input of an STASN model, obtains candidate STASN models with different hyperparameter combinations, obtains an optimal STASN model by minimizing a loss function, and uses the test set to perform icing detection test and result evaluation. The application combines graph machine learning and small sample learning technology, overcomes defects that traditional fan blade icing detection ignores non-Euclidean space structure information of sensors, different icing stage information and excessively relies on a large number of artificially labeled samples, and still has a high icing detection accuracy in an extreme scene with a small sample quantity of sensor data and class imbalance.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of fan equipment fault diagnosis, and more particularly relates to a fan blade icing detection method and system based on a space-time attention twin network. BACKGROUND

[0002] In recent years, renewable power generation, especially wind power generation, has been widely used. The expansion of wind power has caused a large number of wind turbines to be affected by harsh weather conditions. In high humidity and extremely cold weather conditions, the blades of wind turbines are prone to icing. Wind turbine blade icing can cause the unit to shut down, resulting in power loss and mechanical failure of the unit components, thereby affecting the safety, stability and economy of the wind power system.

[0003] Therefore, fan blade icing detection has positive significance. Existing blade icing detection methods include physical mechanism methods and data-driven methods. The physical mechanism method requires the installation of additional sensors, which is less economical. In addition, this method also relies too much on complex professional knowledge. Data-driven methods can overcome the shortcomings of physical mechanism methods and have received more attention. However, existing data-driven methods are still not perfect. These methods cannot consider the contribution of non-Euclidean space structure information and different icing stage information of sensors to icing detection, and rely too much on a large number of manually labeled samples, and cannot better handle class imbalanced samples. SUMMARY

[0004] Traditional data-driven icing detection methods ignore the non-Euclidean space structure information between multiple sensors and the different icing stage information, and cannot better handle small sample and class imbalanced sensor data. The present application provides a fan blade icing detection method and system based on a space-time attention twin network, which uses GAT modules and GRU modules to model the spatial correlation and temporal correlation of sensor data, respectively, to extract spatial features and temporal features. After the spatial features and temporal features are fused, the similarity measurement mechanism of the SN module solves the small sample and class imbalanced icing detection technical problems.

[0005] To achieve the above-mentioned purpose, according to one aspect of the present application, a fan blade icing detection method based on a space-time attention twin network is provided, comprising:

[0006] Collecting multi-sensor data in the SCADA system of a wind farm, pre-processing the multi-sensor data to obtain target sensor data;

[0007] Dividing the target sensor data into a training set, a validation set and a test set;

[0008] According to the similarity information between the target sensor data, an attribute graph is constructed, and the attribute graph is taken as an input of a spatio-temporal attention Siamese network (STASN) model;

[0009] On the basis of artificial experience and grid-based machine automatic search technology, candidate STASN models with different hyperparameter combinations are obtained, and the training set and the validation set are used to train and evaluate the candidate STASN models, respectively;

[0010] By minimizing the loss function, an optimal STASN model is obtained, and the test set is used for icing detection test and result evaluation.

[0011] In some optional embodiments, the STASN model comprises a graph attention network (GAT) module, a gated recurrent unit (GRU) and a Siamese network (SN) module. The attribute graph is subjected to the GAT module and the GRU to extract spatial features and temporal features, respectively, the spatial features and the temporal features are fused to obtain spatio-temporal features, then, the similarity distance between the two spatio-temporal features is calculated by using the similarity measurement mechanism of the SN module, and the probability of the icing state is output by the full connection layer and the softmax layer.

[0012] In some optional embodiments, the preprocessing of the multi-element sensor data to obtain the target sensor data comprises:

[0013] After the data cleaning of the multi-element sensor data, normalization processing is performed. The MI values between the candidate data of the N sensors and the label of the blade icing are calculated, respectively, the MI values are arranged in descending order according to the size, and the top n (n<N) candidate data are selected as the target sensor data.

[0014] In some optional embodiments, the construction of the attribute graph according to the similarity information between the target sensor data comprises:

[0015] Suppose that the n target sensor data are {x 1 ,x 2 ,...,x n}, a weighted undirected graph G=(V,A) is defined as an attribute graph, and the correlation between the target sensor data is modeled as nodes V and A of the graph, V={x 1 ,x 2 ,...,x n}, A∈R n×n , wherein A is a weighted adjacency matrix, and the weight between any two target sensor data x i and x j in the weighted adjacency matrix is described as dist(i,j) is the spatial distance between x i and x j , MI(i,j) is the MI value between x i and x j , σ represents the standard deviation of the spatial distance between n sensors, and ε represents a threshold value.

[0016] According to another aspect of the present application, a fan blade icing detection system based on a spatio-temporal attention twin network is provided, comprising:

[0017] a data acquisition and preprocessing module, configured to acquire multi-sensor data in a SCADA system of a wind farm, and to preprocess the multi-sensor data to obtain target sensor data;

[0018] a data set division module, configured to divide the target sensor data into a training set, a validation set, and a test set;

[0019] an attribute graph construction module, configured to construct an attribute graph according to similarity information between the target sensor data, and to use the attribute graph as an input of a spatio-temporal attention twin network (STASN) model;

[0020] a model training and evaluation module, configured to obtain candidate STASN models with different hyperparameter combinations based on artificial experience and grid-based machine automatic search technology, and to train and evaluate the candidate STASN models using the training set and the validation set, respectively;

[0021] a detection module, configured to obtain an optimal STASN model by minimizing a loss function, and to perform icing detection testing and result evaluation using the test set.

[0022] In some optional embodiments, the STASN model comprises a graph attention network (GAT) module, a gated recurrent unit (GRU), and a twin network (SN) module. The attribute graph is subjected to the GAT module and the GRU to extract spatial features and temporal features, respectively, and the spatial features and the temporal features are fused to obtain spatio-temporal features. Then, the similarity distance between two spatio-temporal features is calculated using a similarity measurement mechanism of the SN module, and the probability of an icing state is outputted through a fully connected layer and a softmax layer.

[0023] In some optional embodiments, the data acquisition and preprocessing module is configured to perform normalization processing after data cleaning of the multi-sensor data. The MI values between candidate data of N sensors and labels of blade icing are calculated, the MI values are arranged in descending order according to their sizes, and the top n (n<N) candidate data are selected as the target sensor data.

[0024] In some optional embodiments, the attribute graph construction module is configured to assume that n target sensor data are {x 1 ,x 2 ,...,x n}, define a weighted undirected graph G=(V,A) as an attribute graph, model the target sensor data and the correlation between the target sensor data as nodes V and A of the graph respectively, V={x 1 ,x 2 ,...,x n}, A∈R n×n , wherein A is a weighted adjacency matrix, and the weight between any two target sensor data x i and x j in the weighted adjacency matrix is described as dist(i,j) is the spatial distance between x i and x j , MI(i,j) is the MI value between x i and x j , and σ represents the standard deviation of the spatial distance between the n sensors, and ε represents a threshold value.

[0025] According to another aspect of the present application, a computer readable storage medium is provided, which stores a computer program, and the computer program is executed by a processor to implement the steps of any of the above methods.

[0026] Overall, compared with the prior art, the above technical solutions conceived by the present application can achieve the following beneficial effects:

[0027] The present application uses GAT and GRU to extract spatial feature information and time feature information of sensor data respectively, so that the icing detection model can not only consider the non-Euclidean spatial structure information between multiple sensors, but also consider the difference information of different icing stages, thereby improving the performance of the icing detection model. In addition, the proposed model is designed based on the SN architecture, which can solve the problems of small sample and class imbalance of sensor data. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 is a flowchart of a method provided by an embodiment of the present application;

[0029] Figure 2 is a whole architecture diagram of an STASN icing detection model provided by an embodiment of the present application;

[0030] Figure 3 is a variable selection result diagram provided by an embodiment of the present application;

[0031] Figure 4The application provides a sparse weighted adjacency matrix result graph of an attribute graph.

[0032] Figure 5 The application provides a structure diagram of an attribute graph. DETAILED DESCRIPTION

[0033] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and should not be used to limit the present application. In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.

[0034] The application provides a fan blade icing detection method based on a spatiotemporal attention Siamese network, as shown in the following steps. Figure 1

[0035] S1: design an overall architecture of a fan blade icing detection model based on a spatiotemporal attention Siamese network (STASN), and the architecture diagram is as shown in the following figure. Figure 2

[0036] The overall architecture includes three parts, namely an attribute graph construction part, a spatiotemporal attention Siamese network (STASN) part and a blade icing detection part. The design idea process of the fan blade icing detection model is as follows: first, an attribute graph is constructed by using multi-sensor data; then, the attribute graph data is subjected to a graph attention network (GAT) module and a gated recurrent unit (GRU) module of a Siamese subnetwork of the STASN to respectively extract spatial features and time features, the spatial features and the time features are fused to obtain spatiotemporal features; finally, a similarity distance between the two spatiotemporal features is calculated by using a similarity measurement mechanism of a Siamese network (SN) module, and a probability of an icing state is output by a full connection layer and a softmax layer.

[0037] ​​Further, the GAT module is used for spatial correlation modeling of multi-sensor data, so as to mine the non-Euclidean space structure information between multi-sensor. Considering the spatial structure of the sensor can provide more useful hidden information for icing detection. The GRU module is used for time correlation modeling of sensor data, so as to mine the time information of different icing stages. Considering the time correlation of sensor data is conducive to distinguishing the difference in contribution of different icing stages to icing detection, such as the phase feature information of the stage close to or in icing state is more important than the phase feature information of the normal state. The metric-based meta-learning method has many advantages, such as: this kind of method has strong generalization and expandability, especially can handle small sample data and class imbalance data. The SN module is a classic metric-based meta-learning method, which has a simpler structure and lower model complexity. The SN module includes two parameter-shared identical twin sub-networks, each of which contains a GAT module and a GRU module. The paired graph data samples are respectively input into the two sub-networks of the STASN model, and the spatial features and time features can be obtained through the GAT module and the GRU module. The two types of features can obtain two new features containing rich spatio-temporal information through feature fusion based on concatenation operation. The similarity distance between the two new fusion features can be used to judge whether the input sample belongs to the same class, i.e. icing class or normal class.

[0038] S2: Collecting multi-sensor data from the Supervisory Control and Data Acquisition (SCADA) system of the wind farm, and pre-processing the sensor data by data cleaning, normalization and variable selection. After data preprocessing, the variable selection result is as shown in Figure 3 According to the mutual information (Mutual Information, MI) value corresponding to the sensor data, the top n (the embodiment of the application preferably 12) sensor data are selected as the target sensor data.

[0039] In the embodiment of the application, step S2 can be realized by the following method:

[0040] The mutual information MI correlation analysis method is used to select the sensor data which contributes more to icing detection, and the specific process includes: first, the MI value between the candidate data of N sensors and the label is calculated; then, the size of the MI value is arranged in descending order, and the top n (n

[0041]

[0042] where H(X) and H(Y) represent the entropy of X and Y respectively, H(X,Y) represents the joint entropy between X and Y, and P(·) represents the probability density function.

[0043] S3: The preprocessed target sensor data is divided into a training set, a validation set and a test set; a sparse weighted adjacency matrix A of the attribute graph is calculated according to the similarity between the target sensor data, a visual result of which is as shown in FIG. 3, and a structure of the attribute graph is as shown in FIG. 4. These attribute graphs become different samples after embedding the target sensor data under different time windows, and any two of the samples are selected for free pairing, and a new label is given according to whether the sample pair belongs to the same class. If the sample pair is of the same class, the label is 1; if the sample pair is of different classes, the label is 0. The above-mentioned sample pair and the corresponding label can be used as the input of the STASN model for supervised learning training. Figure 4 Figure 5

[0044] In the embodiment of the application, the attribute graph is constructed in the following manner:

[0045] Suppose n target sensor data are {x 1 ,x 2 ,...,x n}, a weighted undirected graph G=(V,A) can be defined, and the correlation between the target sensor data and the target sensor data is modeled as the nodes V and A of the graph, i.e. V={x 1 ,x 2 ,...,x n}, A∈R n×n , where A is a weighted adjacency matrix. The correlation between any two target sensor data x i and x j can be calculated by using formula (1), and the MI value between x i and x j is denoted as MI(i,j), which indicates that the greater the value, the greater the correlation between x i and x j in a non-Euclidean space. Here, the spatial distance between x i and x j is defined as dist(i,j), which can be described as:

[0046]

[0047] The Gaussian kernel threshold distance function can be used to construct the sparse weighted adjacency matrix A, where the weight between x i and x j can be described as:

[0048] ​​

[0049] wherein σ represents the standard deviation of the spatial distance between n sensors; ε represents a threshold value, which can be the average distance between n sensors.

[0050] S4: On the basis of artificial experience and grid machine automatic search technology, candidate STASN models with different hyperparameter combinations are obtained, wherein the training set and the validation set are used for training and evaluation of the candidate STASN models, respectively. In order to improve the generalization performance of the STASN model, the training and evaluation process of the candidate STASN model can adopt a K-fold cross-validation strategy.

[0051] S5: An optimal STASN model is obtained by minimizing the loss function, and a test set is used for icing detection test and result evaluation.

[0052] The fan blade icing detection method of the space-time attention twin network. The GAT and GRU are used to extract the spatial feature information and the time feature information of the sensor data, so that the icing detection model can not only consider the non-Euclidean spatial structure information between multiple sensors, but also consider the difference information of different icing stages, thereby improving the performance of the icing detection model. In addition, the proposed STASN model is designed based on the SN architecture, which can solve the problems of small sample and class imbalance of sensor data, improve the precision of fan blade icing detection, and especially the detection result still has good robustness in the extreme scene of small sample and class imbalance.

[0053] It should be noted that, according to the needs of implementation, each step / component described in the present application can be split into more steps / components, or two or more steps / components or part of the operation of the steps / components can be combined into a new step / component, so as to achieve the purpose of the present application.

[0054] Those skilled in the art will readily understand that the above description is only the preferred embodiment of the present application, and is not intended to limit the present application. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for detecting icing on a fan blade based on a spatio-temporal attention twin network, characterized in that, include: Collect multi-sensor data from the SCADA system of the wind farm, preprocess the multi-sensor data to obtain target sensor data; The target sensor data is divided into a training set, a validation set, and a test set; An attribute map is constructed based on the similarity information between target sensor data, and the attribute map is used as the input to the spatiotemporal attention twin network STASN model; Based on human experience and gridded machine automatic search technology, candidate STASN models with different hyperparameter combinations are obtained, and the candidate STASN models are trained and evaluated using training and validation sets respectively. The optimal STASN model is obtained by minimizing the loss function, and the test set is used to test icing detection and evaluate the results. The STASN model includes a graph attention network (GAT) module, a gated recursive unit (GRU), and a Siamese network (SN) module. The attribute graph is processed by the GAT module and the GRU module to extract spatial and temporal features, respectively. The spatial and temporal features are then fused to obtain spatiotemporal features. Then, the similarity measurement mechanism of the Siamese network (SN) module is used to calculate the similarity distance between two spatiotemporal features, and the probability of the icing state is output after passing through a fully connected layer and a softmax layer.

2. The method of claim 1, wherein, The preprocessing of multi-sensor data to obtain target sensor data includes: After data cleaning on the multi-sensor data, normalization processing is performed on the normalized sensor data N The MI value between the candidate data of the sensor and the label of the ice-coated blade is calculated N The MI values are arranged in descending order according to the size, and the top n The candidate data is selected as the target sensor data, wherein n < N.

3. The method of claim 2, wherein, The step of constructing an attribute map based on the similarity information between target sensor data includes: Assumption n The target sensor data is { x 1 , x 2 ,..., x n Define a weighted undirected graph. G = ( V , A As an attribute graph, the relationship between target sensor data and target sensor data is modeled as nodes of the graph. V and A , V ={ x 1 , x 2 ,..., x n }, A ∈ R n×n ,in, A The weighted adjacency matrix contains the sensor data of any two targets. x i and x j The weights between them are described as follows , dist ( i , j ) for x i and x j Spatial distance between them , MI ( i , j )for x i and x j MI value between σ express n The standard deviation of the spatial distance between the sensors ε This represents the threshold.

4. A fan blade icing detection system based on a spatio-temporal attention twin network, characterized in that, include: The data acquisition and preprocessing module is used to acquire multi-sensor data from the SCADA system of the wind farm, preprocess the multi-sensor data, and obtain target sensor data. The dataset partitioning module is used to divide the target sensor data into training set, validation set and test set; The attribute graph construction module is used to construct attribute graphs based on the similarity information between target sensor data, and the attribute graphs are used as input to the spatiotemporal attention twin network STASN model; The model training and evaluation module is used to obtain candidate STASN models with different hyperparameter combinations based on human experience and gridded machine automatic search technology, and to train and evaluate the candidate STASN models using training and validation sets respectively. The detection module is used to obtain the optimal STASN model by minimizing the loss function, and to perform icing detection testing and result evaluation using the test set; The STASN model includes a graph attention network (GAT) module, a gated recursive unit (GRU), and a Siamese network (SN) module. The attribute graph is processed by the GAT module and the GRU module to extract spatial and temporal features, respectively. The spatial and temporal features are then fused to obtain spatiotemporal features. Then, the similarity measurement mechanism of the Siamese network (SN) module is used to calculate the similarity distance between two spatiotemporal features, and the probability of the icing state is output after passing through a fully connected layer and a softmax layer.

5. The system according to claim 4, characterized in that, The data acquisition preprocessing module is configured to perform data cleaning on the multi-element sensor data, perform normalization processing on the data cleaned multi-element sensor data, calculate the mutual information (MI) value between the candidate data of each element sensor and the icing label of the blade, arrange the MI values in descending order, and select the top N candidate data as the target sensor data. N N n n < N.​​​ 6. The system of claim 5, wherein, The attribute graph construction module is used to assume n The target sensor data is { x 1 , x 2 ,..., x n Define a weighted undirected graph. G = ( V , A As an attribute graph, the relationship between target sensor data and target sensor data is modeled as nodes of the graph. V and A , V = { x 1 , x 2 ,..., x n }, A ∈ R n×n ,in, A The weighted adjacency matrix contains the sensor data of any two targets. x i and x j The weights between them are described as follows , dist ( i , j ) for x i and x j Spatial distance between them , MI ( i , j )for x i and x j MI value between σ express n The standard deviation of the spatial distance between the sensors ε This represents the threshold.

7. A computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 3.

Images