Wind turbine anomaly detection method based on deep representation
By employing a depth-based anomaly detection method for wind turbines, and utilizing the dual-anchor loss function and the K-nearest neighbor method, the problem of uneven SCADA data distribution is solved, thereby improving the accuracy of anomaly detection for wind turbines.
Patent Information
- Application Number
- CN202210736346.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-06-27
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2042-06-27
AI Technical Summary
In existing wind turbine monitoring and data acquisition systems, SCADA data is unevenly distributed, and abnormal data is overwhelmed, resulting in weak anomaly detection capabilities and poor fault diagnosis accuracy of intelligent fault diagnosis models.
A depth-based anomaly detection method for wind turbines is adopted. The neural network is trained using a designed dual-anchor loss function to extract depth feature representations. The K-nearest neighbor method is combined for anomaly detection. Features are filtered using mutual information technology. The SMOTE algorithm is used to balance the data. The neural network is designed and the KNN algorithm is used for final detection.
It improves the problem of data imbalance, enhances the performance of outlier identification, and increases the accuracy of anomaly detection.
Smart Images

Figure CN115146718B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of wind turbine anomaly detection, and in particular relates to a wind turbine anomaly detection method based on depth representation. Background Art
[0002] Many intelligent fault diagnosis models have been developed for wind turbine supervisory control and data acquisition (SCADA) systems in order to process large amounts of SCADA data efficiently and accurately. However, there is an overlooked problem in these studies, namely, the SCADA data is unevenly distributed, the abnormal data mining is insufficient, the number of normal data is much larger than the number of abnormal data, and the abnormal state information is easily overwhelmed by the normal state information, which makes these models tend to be biased towards the majority category, and the detection ability of abnormal points is quite weak, resulting in poor accuracy of fault diagnosis. Summary of the invention
[0003] In order to solve the technical problems existing in the prior art, the object of the present invention is to provide a wind turbine anomaly detection method based on depth representation.
[0004] In order to achieve the above purpose and the above technical effect, the technical solution adopted by the present invention is:
[0005] The wind turbine anomaly detection method based on deep representation is mainly used for operating status monitoring of wind turbines. First, the neural network is trained using the designed double anchor loss function to extract the deep feature representation that effectively divides the normal and abnormal areas. Finally, the K nearest neighbor method is used for anomaly detection.
[0006] The wind turbine anomaly detection method based on deep representation specifically includes the following steps:
[0007] Step 1: Extract SCADA data variables, including wind speed, wind direction angle, active power, wind turbine speed, pitch angle, pitch angle DC, cabin temperature, and pitch motor temperature;
[0008] Step 2: Label the abnormal data and normal data, and resample the abnormal data;
[0009] Step 3: Design a neural network;
[0010] Step 4: Design a new dual-anchor loss function, use the designed dual-anchor loss function to train the neural network, and extract the deep feature representation that effectively divides the normal and abnormal areas;
[0011] Step 5: Use K nearest neighbor method to detect wind turbine anomalies.
[0012] In step 1, the steps of extracting SCADA data variables include:
[0013] The mutual information technology is used to perform feature selection on the multidimensional variables of SCADA data, and the variables most relevant to the abnormal state of the generator are selected as the features for subsequent anomaly detection;
[0014] The mutual information (MI) of two discrete events X and Y is defined as:
[0015]
[0016] Among them, p(x,y) represents the joint probability distribution of X and Y, p1(x) represents the independent probability distribution of X, and p2(y) represents the independent probability distribution of Y;
[0017] In the wind power system, let the input characteristic be Xm and the output characteristic be Y. The input characteristic and the output characteristic are processed as follows:
[0018] Step 1: Sort the input feature Xm and output feature Y from small to large;
[0019] Step 2: assign the largest third to 2, the smallest third to 0, and the middle third to 1. After this processing, each input feature and output feature has only 3 possible values, and its independent probability distribution and joint probability distribution can be easily found;
[0020] Step 3: After step 2, the independent probability distribution of each feature's possible values is 1 / 3, and the joint probability distribution between the input and output features is as follows:
[0021]
[0022]
[0023] Where i, j = 0, 1, 2; L is the total number of samples; l is the lth sample;
[0024] Step 4: Calculate the mutual information of input and output features. The calculation formula of the processed mutual information MI is:
[0025]
[0026] This greatly reduces the computational dimension of the sample, and the mutual information between the input and output can be calculated. Formula (4) is used to screen the SCADA high-dimensional variable features, and the correlation threshold is set to 0.05. The feature set below the threshold will be removed. After mutual information processing, the extracted features include: wind speed, wind direction angle, active power, wind turbine speed, pitch angle, pitch angle DC, cabin temperature, and pitch motor temperature.
[0027] In step 2, the steps of labeling abnormal data and normal data and resampling abnormal data include:
[0028] First, a training set is constructed, and the selected features are used as the input of the neural network model. The output of the neural network model needs to be labeled. The abnormal values are labeled according to the alarm and fault information of the SCADA data, and the rest are normal values. Since the total amount of abnormal data is much smaller than the total amount of normal data, the SMOTE algorithm is used to perform linear interpolation on the abnormal signal to obtain a balance between normal data samples and abnormal samples; the SMOTE data resampling method is as follows:
[0029] x gen =x i +rand(0,1)(x i -x ik ) (5)
[0030] Among them, x gen is the generated sample, x i is the original data point, x ik For x i The kth neighbor point of , rand(0,1)∈[0,1] is a random number.
[0031] In step 3, the steps of designing the neural network include:
[0032] First, construct a training method, and input the following types of data into the neural network in sequence: abnormal features, normal features, abnormal anchor features, and normal anchor features. Each type is obtained by step one and step two. Each time the above four types of data are input, a gradient descent is performed through the constructed loss function, and it is continuously iterated. Finally, when the loss reaches the minimum threshold, the iteration is stopped, and the neural network model training is completed. In order to obtain different types of anchor data, Kmean clustering is performed on the normal and abnormal classes respectively, and the respective cluster centers are used as anchor data.
[0033] After that, the data is input and the convolutional neural network is used to extract local information between the data. Finally, the data is output through the fully connected layer and the feature is extracted using 3*1 convolution. During training, the double anchor loss function is used for deep feature extraction. During the training process, the neural network output layer of the model is z-score normalized to maintain the boundedness and stability of the data output.
[0034] In step 4, the double anchor loss function is as follows:
[0035]
[0036] Among them, g θ (·) is the feature mapping function, are the i-th positive sample and negative sample respectively, x a-n ,x a-p are negative sample anchors and positive sample anchors respectively, and α is the interval threshold between positive and negative anchors.
[0037] In step five, the trained model is used as a deep feature extractor for the original SCADA variables. Finally, the KNN algorithm is used to detect anomalies in the data, that is, the original deep features are stored. When new data appears, the new data is feature mapped through the deep feature extractor, and the obtained mapping features are input into KNN to determine whether there is an abnormality in the wind turbine.
[0038] Compared with the prior art, the present invention has the following beneficial effects:
[0039] The deep feature representation of SCADA data is performed by using neural network combined with the designed double anchor loss function, which improves the imbalance of data, is beneficial to improving the recognition performance of outliers, and is conducive to anomaly detection. BRIEF DESCRIPTION OF THE DRAWINGS
[0040] Figure 1 It is the basic topological structure diagram of the SCADA convolutional neural network of the present invention;
[0041] Figure 2 It is a schematic diagram of the principle of the double anchor loss function of the present invention. DETAILED DESCRIPTION
[0042] The present invention is described in detail below so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby making a clearer and more definite definition of the protection scope of the present invention.
[0043] A brief summary of one or more aspects is given below to provide a basic understanding of these aspects. This summary is not an exhaustive overview of all conceived aspects, and is neither intended to identify the key or critical elements of all aspects nor to define the scope of any or all aspects. Its only purpose is to give some concepts of one or more aspects in a simplified form as a prelude to a more detailed description that will be given later.
[0044] like Figure 1-2 As shown in the figure, the wind turbine anomaly detection method based on deep representation is mainly used for the operation status monitoring of wind turbines. First, the neural network is trained using the designed double anchor loss function to extract the deep feature representation that effectively divides the normal and abnormal areas. Finally, the K nearest neighbor method is used for anomaly detection, which specifically includes the following steps:
[0045] Step 1: Extract SCADA data variables, including wind speed, wind direction angle, active power, wind turbine speed, pitch angle, pitch angle DC, cabin temperature, pitch motor temperature, etc.
[0046] In step 1, mutual information technology is used to perform feature selection on the multidimensional variables of SCADA data, and the variables most relevant to the abnormal state of the generator are selected as the features for subsequent anomaly detection.
[0047] Mutual information (MI) is a useful information measure in information theory, which is used to measure the correlation between two sets of events. The mutual information (MI) of two discrete events X and Y is defined as:
[0048]
[0049] Among them, p(x,y) represents the joint probability distribution of X and Y, p1(x) represents the independent probability distribution of X, and p2(y) represents the independent probability distribution of Y.
[0050] It can be seen from formula (1) that to calculate the mutual information between two variables, the independent probability distribution and joint probability distribution of each variable must be calculated. In the wind power system, let the input feature be Xm and the output feature be Y. Since the dimension of the candidate input feature is particularly large, its dimension is close to the number of samples, the mutual information between the input and output features cannot be calculated directly using the MI algorithm, so the input and output features are processed as follows:
[0051] Step 1: Sort the input feature Xm and output feature Y from small to large;
[0052] Step 2: assign the largest third to 2, the smallest third to 0, and the middle third to 1. After this processing, each input and output feature has only 3 possible values, and its independent probability distribution and joint probability distribution can be easily found;
[0053] Step 3: After step 2, the independent probability distribution of each feature's possible values is 1 / 3, and the joint probability distribution between the input and output features is as follows:
[0054]
[0055]
[0056] Where i, j = 0, 1, 2; L is the total number of samples; l is the lth sample;
[0057] Step 4: Calculate the mutual information of input and output features. Based on the above independent and joint probability distributions, the calculation formula of the processed mutual information MI can be written as follows:
[0058]
[0059] This greatly reduces the computational dimension of the sample, and the mutual information between the input and output can be calculated. Formula (4) is used to screen the SCADA high-dimensional variable features, and the correlation threshold is set to 0.05. The feature set below the threshold will be removed. After mutual information processing, the extracted features include: wind speed, wind direction angle, active power, wind turbine speed, pitch angle, pitch angle DC, cabin temperature, variable pitch motor temperature, etc.
[0060] Step 2: Label the abnormal data and normal data, and resample the abnormal data.
[0061] In step 2, first construct a training set and use the selected features as the input of the neural network model. The output of the model needs to be labeled. The abnormal values are labeled according to the alarm and fault information of the SCADA data, and the rest are normal values. Since the total amount of abnormal data is much smaller than the total amount of normal data, the SMOTE algorithm is used to perform linear interpolation on the abnormal signal to obtain a balance between normal data samples and abnormal samples. The SMOTE data resampling method is as follows:
[0062] x gen =x i +rand(0,1)(x i -x ik ) (5)
[0063] Among them, x gen is the generated sample, x i is the original data point, x ik For x i The kth neighbor point of , rand(0,1)∈[0,1] is a random number.
[0064] Step 3: Design a neural network
[0065] First, construct a training method and input the following types of data into the network in turn: abnormal features, normal features, abnormal anchor features, and normal anchor features. Each type is obtained by steps one and two. Each time the above four types of data are input, a gradient descent is performed through the constructed loss function. Continue to iterate, and finally stop the iteration when the loss reaches the minimum threshold, and the model training is completed. In order to obtain different categories of anchor data, Kmean clustering is performed on the normal and abnormal classes respectively, and the respective cluster centers are used as anchor data;
[0066] After that, the data is input, and the local information between the data is extracted using the convolutional neural network, and finally output through the fully connected layer. The network topology is as follows Figure 1As shown in the figure, 3*1 convolution is used for feature extraction, and the double anchor loss function is used for deep feature extraction during training. During the training process, the output layer of the neural network of the model is z-score normalized to maintain the boundedness and stability of the data output.
[0067] Step 4: Based on the triplet loss function, a new double anchor loss function is designed to realize the abnormal feature extraction of wind turbines, such as Figure 2 As shown in Figure 2. Triplet loss is a loss function for deep learning, which is mainly used to train samples with small differences, such as faces, fine-grained classification, etc. The designed double anchor loss function is as follows:
[0068]
[0069] Among them, g θ (·) is the feature mapping function, are the i-th positive sample and negative sample respectively, x a-n ,x a-p are negative sample anchors and positive sample anchors respectively, and α is the interval threshold between positive and negative anchors.
[0070] Step 5: The trained model is used as a deep feature extractor for the original SCADA variables. Finally, the KNN algorithm is used to detect anomalies in the data, that is, the original deep features are stored. When new data appears, the new data is mapped through the deep feature extractor, and the mapping features are input into the KNN to determine whether the wind turbine has an abnormality.
[0071] Parts or structures not specifically described in the present invention may adopt existing technologies or existing products and will not be described in detail here.
[0072] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.
Claims
1. Wind turbine anomaly detection method based on deep representation, It is characterized in that Firstly, the designed dual anchor loss function is used to train the neural network to extract the deep feature representation that effectively divides the normal and abnormal regions. Finally, the K nearest neighbor method is used for anomaly detection. The detection method comprises the following steps: Step 1: Extract SCADA data variables, including wind speed, wind direction angle, active power, wind turbine speed, pitch angle, pitch angle DC, cabin temperature, and pitch motor temperature; Step 2: Label the abnormal data and normal data, and resample the abnormal data; Step 3: Design a neural network; Step 4: Design a new dual-anchor loss function, use the designed dual-anchor loss function to train the neural network, and extract the deep feature representation that effectively divides the normal and abnormal areas; Step 5: Use K nearest neighbor method to detect wind turbine anomalies; In step 3, the steps of designing the neural network include: First, construct a training method, and input the following types of data into the neural network in sequence: abnormal features, normal features, abnormal anchor features, and normal anchor features. Each type is obtained by step one and step two. Each time the above four types of data are input, a gradient descent is performed through the constructed loss function, and it is continuously iterated. Finally, when the loss reaches the minimum threshold, the iteration is stopped, and the neural network model training is completed. In order to obtain different types of anchor data, Kmean clustering is performed on the normal and abnormal classes respectively, and the respective cluster centers are used as anchor data. After that, the data is input, and the local information between the data is extracted using the convolutional neural network. Finally, the data is output through the fully connected layer, and the feature is extracted using 3*1 convolution. During training, the double anchor loss function is used for deep feature extraction. During the training process, the output layer of the neural network of the model is z-score normalized to maintain the boundedness and stability of the data output. In step 4, the double anchor loss function is as follows: Among them, g θ (·) is the feature mapping function, are the i-th positive sample and negative sample respectively, x a-n ,x a-p They are negative sample anchors and positive sample anchors, respectively. Kmean clustering is performed on the positive samples and negative samples respectively, and their respective cluster centers are used as the positive sample anchors and negative sample anchors respectively. α is the interval threshold between the positive and negative anchors.
2. The method for detecting abnormality of a wind turbine generator system based on deep representation according to claim 1, It is characterized in that In step 1, the steps of extracting SCADA data variables include: The mutual information technology is used to perform feature selection on the multidimensional variables of SCADA data, and the variables most relevant to the abnormal state of the generator are selected as the features for subsequent anomaly detection; The mutual information MI of two discrete events X and Y is defined as: Among them, I represents mutual information MI, p(x,y) represents the joint probability distribution of X and Y, and p 1 (x) represents the independent probability distribution of X, p 2 (y) represents the independent probability distribution of Y; In the wind power system, let the input characteristic be Xm and the output characteristic be Y. The input characteristic and the output characteristic are processed as follows: Step 1: Sort the input feature Xm and output feature Y from small to large; Step 2: assign the largest third to 2, the smallest third to 0, and the middle third to 1. After this processing, each input feature and output feature has only 3 possible values, and its independent probability distribution and joint probability distribution can be easily found. Step 3: After step 2, the independent probability distribution of each feature's possible values is 1 / 3, and the joint probability distribution between the input and output features is as follows: Where i, j = 0, 1, 2; L is the total number of samples; l is the lth sample; X m,i represents the input feature with only three possible values; Y j Indicates that the above output features have only 3 possible values; Step 4: Calculate the mutual information of input and output features. The calculation formula of the processed mutual information MI is: This greatly reduces the computational dimension of the sample, and the mutual information between the input and output can be calculated. Formula (4) is used to screen the SCADA high-dimensional variable features, and the correlation threshold is set to 0.
05. The feature set below the threshold will be removed. After mutual information processing, the extracted features include: wind speed, wind direction angle, active power, wind turbine speed, pitch angle, pitch angle DC, cabin temperature, and variable pitch motor temperature.
3. The wind turbine anomaly detection method based on deep representation according to claim 1, It is characterized in that In step 2, the steps of labeling abnormal data and normal data and resampling abnormal data include: First, a training set is constructed, and the selected features are used as the input of the neural network model. The output of the neural network model needs to be labeled. The abnormal values are labeled according to the alarm and fault information of the SCADA data, and the rest are normal values. Since the total amount of abnormal data is much smaller than the total amount of normal data, the SMOTE algorithm is used to perform linear interpolation on the abnormal signal to obtain a balance between normal data samples and abnormal samples; the SMOTE data resampling method is as follows: x gen =x i +rand(0,1)(x i -x ik ) (5) Among them, x gen is the generated sample, x i is the original data point, x ik For x i The kth neighbor point of , rand(0,1)∈[0,1] is a random number.
4. The method for detecting abnormality of a wind turbine generator system based on deep representation according to claim 1, It is characterized in that In step five, the trained model is used as a deep feature extractor for the original SCADA variables. Finally, the KNN algorithm is used to detect anomalies in the data, that is, the original deep features are stored. When new data appears, the new data is feature mapped through the deep feature extractor, and the obtained mapping features are input into KNN to determine whether there is an abnormality in the wind turbine.
