Wind turbine generator abnormal state detection method based on KAform model

Abstract

The invention discloses a wind turbine generator abnormal state detection method based on a KAform model. The method comprises the following steps: a, data preprocessing; b, construction of a detection model: on the basis of the original ViT, replacing an encoder of the original ViT with a KAN from a simple MLP, constructing a KAformer model, and adding a preposed LSTM network; c, training the model; d, testing the model: calculating an early warning threshold value and an alarm threshold value; constructing an abnormal state judgment index, and judging an abnormal operation state occurrence time period of the wind turbine generator according to the judgment index; and e, detecting the abnormal state of the wind turbine generator by using the detection model which is tested to be qualified. According to the method, the detection model capable of well establishing the normal operation state logic of the wind turbine generator is obtained by improving the original ViT model, then the abnormal state judgment index is established through the KLD calculation and the KDE method, the abnormal operation state of the wind turbine generator can be accurately detected, an alarm is given in time, and safe operation of the wind turbine generator is guaranteed.

Description

Technical Field

[0001] The present invention relates to a method for detecting abnormal states of a wind turbine, belonging to the technical field of detection. Background Art

[0002] Wind power generation is a key component of sustainable energy generation, generating clean electricity from wind, without producing harmful emissions and without relying on finite resources. However, wind turbines are usually installed in remote geographical locations, and the harsh working environment significantly increases the labor and material costs required to maintain the safe operation of wind turbines. Therefore, how to conduct early fault monitoring and warning of wind turbines is of great significance for improving the operation and maintenance efficiency of wind farms and maintaining the safe operation of the turbines.

[0003] Currently, the state monitoring methods of wind turbines mainly include signal trend analysis, physical model-based analysis methods, and data-driven methods. Among them, the signal trend analysis method usually requires installing a large number of sensors to collect various signals, involving relatively high technical costs. The physical model-based method requires researchers to have a deep understanding of the physical characteristics of each component of the wind turbine to create an accurate physical model, which is often very difficult. The data-driven method relies on a large amount of high-quality data, and the data collected by the Supervisory Control and Data Acquisition (SCADA) system equipped in the wind turbine can just meet this requirement. The feature learning and non-linear expression ability of the neural network can extract the high-dimensional characteristics of the SCADA data and can effectively mine the key information in the abnormal state.

[0004] The invention patent with the publication number CN109740175A in Chinese patents discloses a method for detecting outliers in the power curve data of a wind turbine. Based on the real-time operation data of the Supervisory Control and Data Acquisition (SCADA) system of the wind turbine, including wind speed, active power, etc., after a series of preprocessing steps, the data is divided according to certain wind speed and power intervals respectively; further, three outlier detection algorithms, namely Average Distance to the Cluster (AVDC), Local Outlier Factor (LOF), and Density-Based Spatial Clustering of Applications with Noise (DBSCAN), are used to detect suspected outliers; finally, the real outliers are identified from the suspected outliers based on the real outlier discrimination criterion. This method is data-driven and has no special requirements for other information of the wind turbine, and has strong universality. While combining the advantages of the mainstream outlier detection methods, it takes into account the characteristics of the power curve data set and provides a guarantee for the data quality, and has strong theoretical and application value. However, the reliability and accuracy of the existing data-driven detection methods are still not ideal enough, and it is necessary to improve them. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for detecting abnormal states of wind turbine units based on the KAformer model in view of the drawbacks of the prior art, so as to improve the reliability and accuracy of detecting abnormal states of wind turbine units.

[0006] The problems of the present invention are solved by the following technical solutions:

[0007] A method for detecting abnormal states of wind turbine units based on the KAformer model, the method comprising the following steps:

[0008] a. Data preprocessing: Perform data cleaning and normalization on the original data of the operation of the wind turbine unit obtained from the SCADA system, select target parameters and input parameters, and divide the training set and the test set;

[0009] b. Construction of the detection model: On the basis of the original ViT, replace its encoder with a KAN from a simple MLP, construct a KAformer model, and add a pre-LSTM network to the front of the KAformer model to obtain the required detection model;

[0010] c. Training of the model: Input the training set data into the detection model to train the detection model;

[0011] d. Testing of the model: Input the test set data into the trained detection model to obtain the predicted values of the target parameters; Calculate the KLD between the predicted values of the target parameters and the true values to obtain the state monitoring curve, and then use the KDE method to calculate the early warning threshold and the alarm threshold; Construct an abnormal state evaluation index according to the state monitoring curve, the early warning threshold and the alarm threshold, and judge the time period of the abnormal operation state of the wind turbine unit according to the evaluation index, so as to realize the testing of the model;

[0012] e. Detecting the abnormal state of the wind turbine unit by using the detection model qualified in the test: Input the operation data of the wind turbine unit to be measured into the detection model qualified in the test to obtain the state monitoring curve, and judge the time period of the abnormal operation state of the wind turbine unit according to the evaluation index.

[0013] For the above method for detecting abnormal states of wind turbine units based on the KAformer model, the specific process of the data preprocessing is as follows:

[0014] a1. Data cleaning and normalization: The specific steps of using the quartile method for data cleaning are as follows: Arrange the values of the curve in ascending order into four intervals according to power and wind speed, and the power in each wind speed interval is distributed in a quartile range in ascending order, and calculate the interquartile range (IQR, the difference between the third quartile Q 3 and the first quartile Q 1 ), and then calculate the upper limit value W′ uand the lower limit value W' l . The formula is as follows:

[0015] IQR = Q 3 -Q 1

[0016]

[0017] Data higher than the upper limit value W' u or lower limit value W' l is determined as an outlier and removed.

[0018] SCADA data includes parameter data in various aspects such as wind speed, temperature, power, voltage, and current. It is necessary to normalize the cleaned data and convert various variables into the range of [0, 1]. The specific formula is as follows:

[0019]

[0020] a2. Selection of target parameters and input parameters: Select the parameters corresponding to the actual fault as the target parameters, and select the parameters with a Spearman correlation coefficient greater than 0.5 with the target parameters as the input parameters;

[0021] a3. Division of training set and test set: Divide the data during the period when the wind turbine is considered to be operating normally into the training set, and divide the data in the entire monitoring interval into the test set;

[0022] For the above-mentioned wind turbine abnormal state detection method based on the KAformer model, during the construction process of the detection model, the construction methods of each sub-layer are as follows:

[0023] b1. Pre-LSTM layer: Set the number of layers of the pre-LSTM layer to 1, and the number of neurons to 64;

[0024] b2. KAformer model: The KAformer model includes an embedding layer, an encoder, and a decoder. The embedding layer is composed of a convolutional layer and a fully connected layer, and the input channel is set to 11; the encoder is composed of L encoding layers stacked in sequence, and each encoding layer is composed of a self-attention mechanism layer, a residual connection and normalization layer, and a feed-forward network layer. Set the MLP dimension to 512, the number of hidden layers to 256, and the multi-head attention mechanism to 8 heads; the decoder is composed of KANlinear, and the number of nodes N is set to 10, and the spline parameter is 3.

[0025] For the above-mentioned wind turbine abnormal state detection method based on the KAformer model, the training of the model consists of 200 epochs, the batch size is 256, the learning rate is 0.0001, dropout is 0.1, the activation function is GELU, and the optimizer uses Adam.

[0026] For the above-mentioned abnormal state detection method of wind turbines based on the KAformer model, the calculation methods of the warning threshold and the alarm threshold are as follows:

[0027] Input the test set data into the trained detection model to obtain the predicted values of the target parameters, and perform KLD calculation on them with the true values:

[0028]

[0029] Among them, P(x) represents the probability distribution of the predicted values, Q(x) represents the probability distribution of the true values, KLD(P||Q) represents the KLD from P(x) to Q(x), and the KLDs of the data sample points form a detection curve characterizing the operating state of the wind turbine; Take part of the training set data, set a sliding time window with a window of 24 hours and a step size of 1 minute for sampling, and use the KDE method to calculate the warning threshold and the alarm threshold. The specific formula is:

[0030]

[0031] In the formula, p(h) represents the density estimate value at point h, h k is the k-th sample data point, N is the number of samples included in the set sliding time window, σ is the kernel function bandwidth coefficient, K(·) is the kernel function, and the safety confidence level s c is set, then the warning threshold T 1 is calculated by the following formula:

[0032]

[0033] The alarm threshold T 2 takes the maximum value of p(h) of part of the training set data.

[0034] For the above-mentioned abnormal state detection method of wind turbines based on the KAformer model, the safety confidence level s c is 98%.

[0035] For the above-mentioned abnormal state detection method of wind turbines based on the KAformer model, the kernel function K(·) is a Gaussian kernel function, and its expression is as follows:

[0036]

[0037] The above-mentioned abnormal state detection method for wind turbines based on the KAformer model, and the evaluation indexes for judging the abnormal state include: the number of early warnings, early warning sample points, early warning duration, alarm sample points, alarm duration, and the first alarm date. The part where the detection curve crosses the early warning threshold but does not cross the alarm threshold is regarded as the early warning state. If the duration for which the monitoring curve exceeds the alarm threshold exceeds 60 minutes, it is regarded that the wind turbine is in an abnormal operating state, and the date corresponding to the sample point where the alarm threshold is first crossed is the first alarm date.

[0038] In the present invention, the original ViT model is improved by KAN to obtain the KAformer model with strong data deep feature extraction ability, which can well establish the normal operating state logic of wind turbines. Then, the abnormal state evaluation indexes are constructed through KLD calculation and KDE method, which can accurately detect the abnormal operating state of wind turbines, give an alarm in time, and ensure the safe operation of wind turbines. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The present invention will be further described in detail below with reference to the accompanying drawings.

[0040] Figure 1 is the overall flowchart of the present invention;

[0041] Figure 2 is the time domain diagram of each parameter;

[0042] Figure 3 is the structural diagram of the detection model;

[0043] Figure 4 is the state detection result diagram of the method designed by the present invention for example data.

[0044] Each symbol in the text is: N t is the length of the time series, d model is the channel code, d k is the dimension of the key vector; W i Q , W i K , W i V are the query transformation matrix, key transformation matrix, and value transformation matrix respectively. Q is the query matrix, K is the key matrix, and V is the value matrix; Z i is the self-attention of the i-th head, W O is the linear transformation matrix b 1 , b 2 , W 1 , W 2 are the biases and weights of two linear transformation layers respectively. LN is the normalization layer; sublayer is the multi-head self-attention or fully connected feed-forward network, is the output of the multi-head attention module, is the output of the feedforward neural network, φ p,m is a continuous univariate function, where n, p, and q represent the number of network nodes respectively. B k (x) is a spline function, P(x) represents the probability distribution of the predicted value, Q(x) represents the probability distribution of the true value, h = {h 1 , h 2 , …, h N} is the KLD value of the sample points, N is the number of samples included in the set sliding time window, σ is the kernel function bandwidth coefficient, K(·) is the kernel function, and T 1 is the calculated warning threshold, and T 2 is the alarm threshold. Specific implementation manner

[0045] In view of the drawbacks of the prior art, the present invention designs a method for detecting abnormal states of wind turbines based on the KAformer model. On the basis of the original Vision Transformer (ViT), the decoder thereof is replaced by a Kolmogorov - Arnold Networks (KAN) from a simple MLP, a KAformer model is constructed, and combined with a specific threshold setting method, it can perform abnormal state warning more accurately.

[0046] The steps for detecting abnormal states of the present invention are as follows:

[0047] S1. Data preprocessing: The original data obtained from the SCADA system cannot be directly used for state detection. The preprocessing steps for it include: data cleaning and normalization, selection of target parameters and input parameters, division of training set and test set, etc.;

[0048] S2. Model construction: On the basis of the original ViT, the encoder thereof is replaced by a KAN from a simple MLP, a KAformer model is constructed, and a pre - placed LSTM network is added in the front of the KAformer model to obtain the required detection model. The data with time features extracted by the pre - placed LSTM network will be further processed in the KAformer model;

[0049] S3. Input the training set data into the constructed detection model for training, adjust the hyperparameters, establish a complex relationship between the input parameters and the target parameters, and obtain a network model representing the normal operation state logic of the wind turbine;

[0050] S4. Testing of the model: Input the test set data into the trained detection model to obtain the predicted values of the target parameters; calculate the Kullback-Leibler divergence (KLD) between the predicted values and the true values of the target parameters to obtain the KLD of all sample points in the data set; plot with the sampling time of the sample points as the abscissa and the magnitude of the KLD of the sample points as the ordinate. The curve formed by the KLD of this series of points is the state monitoring curve, and then use the kernel probability density estimation (KDE) method to obtain the warning threshold and the alarm threshold; construct an abnormal state evaluation index based on the state monitoring curve, the warning threshold and the alarm threshold, and judge the time period of the abnormal operation state of the wind turbine according to the evaluation index, so as to realize the testing of the model;

[0051] S5. Detect the abnormal state of the wind turbine using the tested detection model: Input the operation data of the wind turbine to be measured into the tested detection model to obtain the state monitoring curve, and judge the time period of the abnormal operation state of the wind turbine according to the evaluation index.

[0052] In step S1, the specific process of data preprocessing is as follows:

[0053] S11. Data cleaning and normalization: The specific steps of using the quartile method for data cleaning are as follows: Arrange the values of the curve into four intervals from small to large according to power and wind speed. The power in each wind speed interval is distributed from small to large within a quartile range, and calculate the interquartile range (IQR, the difference between the third quartile Q 3 and the first quartile Q 1 ), and then calculate the upper limit value W′ u and the lower limit value W′ l . The formula is as follows:

[0054] IQR = Q 3 -Q 1

[0055]

[0056] Data higher than the upper limit value W′ u or lower than the lower limit value W′ l is determined as an outlier and removed.

[0057] The SCADA data includes parameter data in various aspects such as wind speed, temperature, power, voltage and current. It is necessary to normalize the cleaned data and convert various variables into the range of [0, 1]. The specific formula is as follows:

[0058]

[0059] S12. Selection of target parameters and input parameters: Select the parameters corresponding to the actual fault as the target parameters, and select the parameters with strong correlation with the target parameters as the input parameters through Spearman correlation analysis;

[0060] S13. Dataset division: Divide the data during the time period when the wind turbine is considered to be operating normally into the training set, and divide the data in the entire monitoring interval into the test set;

[0061] In step S2, the specific process of building each sub-layer of the detection model is as follows:

[0062] S21. Pre-LSTM layer: The data performs initial time feature extraction in this layer. Set its number of layers to 1 layer and the number of neurons to 64;

[0063] S22. KAformer model: This model is mainly composed of an embedding layer, an encoder, and a decoder. Among them: The embedding layer is composed of a convolutional layer and a fully connected layer, which transforms the input data into an embedding vector, and the input channel is set to 11; The encoder is composed of L encoding layers stacked in sequence. Each encoding layer consists of a self-attention mechanism layer, a residual connection and normalization layer, and a feed-forward network layer. Set the MLP dimension to 512, the number of hidden layers to 256, and the multi-head attention mechanism to 8 heads; The decoder is composed of KANlinear, and the number of nodes N is set to 10, and the spline parameter is 3, indicating the use of cubic spline curves.

[0064] In step S3, the model training consists of 200 epochs, the batch size is 256, the learning rate is 0.0001, dropout is 0.1, the activation function is RELU, and the optimizer uses Adam.

[0065] In step S4, the specific process of the threshold setting method is as follows:

[0066] Input the test set data into the trained model to obtain the predicted values of the target parameters, calculate the KLD with the true values to obtain the monitoring curve representing the operating state of the wind turbine; Then use the KDE method, adopt the KLD of part of the training set data and set a sliding time window to calculate the warning threshold and the alarm threshold;

[0067] In step S4, the evaluation indicators obtained in the abnormal state detection include: the number of warnings, warning sample points, warning duration, alarm sample points, alarm duration, and the first alarm date, etc.; the warning state is the part where the detection curve crosses the warning threshold but does not cross the alarm threshold. The number of times exceeding the warning threshold in this state is the number of warnings; in one warning state, the start and end sample points can be determined, and the time interval between the two sample points can determine the warning duration; the alarm state is the part where the detection curve crosses the alarm threshold. In this state, the alarm sample points can be determined, and then the alarm duration can be determined; the first alarm date is the time date corresponding to the sample point that first exceeds the alarm threshold. Figure 4 The abscissa represents the sample points of SCADA data, and each sample point has a corresponding actual sampling time. Therefore, the order of the sample points on the abscissa represents the front and back of the time axis; the sampling interval of each sample point is 1 minute. As described above, determining the interval between the start and end sample points can determine the warning duration and the alarm duration; as Figure 4 shown: The start and end sample points of the warning state are the 71000th and the 71681st respectively. The sampling interval of each sample point is 1 minute. So the warning time is 681 minutes, as shown in the data in the following table; similarly, the alarm duration is determined in this way. The first alarm date is the actual sampling date corresponding to the sample point that first crosses the alarm threshold - the 84243rd sample point, which is April 29, 2015.

[0068] The specific abnormal state detection methods for steps S4 and S5 are: If the duration for which the monitoring curve exceeds the alarm threshold exceeds 60 minutes, it is considered that the wind turbine is in an abnormal operating state, and the date corresponding to the sample point that first crosses the alarm threshold is the first alarm date.

[0069] The present invention will be described in detail below with reference to examples. The specific processes of each step are as follows:

[0070] 1) The specific process of data cleaning and normalization is as follows: First, the original data set is screened to remove data points with a wind speed of 0, a power of 0, and wind speed values outside the range of the cut-in wind speed (4 m / s) and the cut-out wind speed (25 m / s). Subsequently, the quartile method is used to remove the outliers in the data set. Finally, the cleaned data needs to be normalized to convert various variables into the range of [0, 1].

[0071] The SCADA data used in the present invention represents gearbox failures and contains 182852 samples; after preprocessing, the data set remains 169551 samples.

[0072] 2) The specific process of target parameter and input parameter selection is as follows: The target parameter is selected as the bearing temperature of the gearbox; the input parameters are selected through Spearman correlation analysis to be the parameters with a correlation coefficient greater than 0.5 with the target parameter, and finally 12 parameters are selected. The time-domain diagrams of the obtained parameters are as shown in Figure 2 shown, and the names of the parameters are as shown in the following table.

[0073]

[0074] 3) The specific process of training set and test set division is as follows: The data set had faults such as gear tooth surface wear and tooth root fracture in the gearbox on July 14, 2015, and after repair, it was restarted in September of the same year. The data time period was from January 1 to September 30, 2015, for a total of 219 days. Among them, the data collected from January 1 to April 1 (from 0 to the 64157th sample point) was considered to be able to characterize the normal operation state of the wind turbine generator set, and it was divided into the training set, while the entire data set was used as the test set in the prediction stage.

[0075] 4) The specific process of model construction is as follows: The KAformer model consists of the following parts in sequence: the embedding layer, the encoder, and the decoder.

[0076] After passing through the LSTM layer, the data is preprocessed into (where N t is the time series length, and d model is the channel code) time series data matrix. The embedding layer performs data segmentation and flattening on this matrix to form multiple one-dimensional vectors, and then becomes embedding vectors through linear transformation. The embedding layer generally realizes the above operations through convolutional layers and fully connected layers, and the mathematical formulas before and after are as follows:

[0077]

[0078] In the formula represents the embedding vector output by the embedding layer, E pos represents the position encoding vector, x class represents the token [class], which is the position of '0', and x′ i (i = 1, 2,..., N t ) represents the channel vector at the i-th position.

[0079] The encoder is composed of L encoding layers stacked in sequence. First, in one encoding layer, the data will pass through the Norm layer and be divided into three parts, Q, K, and V, and input into the multi-head attention module for self-attention calculation, and a scaling factor Calculate the relative weights with the softmax function to obtain the contribution degree of this sequence. In the present invention, the multi-head attention takes eight heads, and the eight sequences are spliced together through a fully connected layer. Then, a feed-forward neural network is connected after the multi-head attention mechanism, and the GELU activation function is added to enhance the parts with larger feature contributions and suppress the smaller parts. Finally, each sub-encoding layer is equipped with a residual connection and a normalization layer to ensure the convergence of the model. The total process of stacking and calculating through L encoding layers can be successively represented by the following formulas:

[0080]

[0081] Z i = attention(QW i Q ,KW i K ,VW i V )

[0082] MultiHead(Q,K,V) = Concat(Z 1 ,...,Z h )W O

[0083] Here, d k —— The dimension of the key vector; W i Q , W i K , W i V —— Query transformation matrix, key transformation matrix and value transformation matrix, respectively responsible for converting the input vector into query matrix Q, key matrix K and value matrix V; Z i —— Self-attention of the i-th head; W O —— Linear transformation matrix.

[0084] FFN(x) = max(0,xW 1 +b 1 )W 2 +b 2

[0085] Here, b 1 , b 2 , W 1 , W 2 are the biases and weights of the two linear transformation layers respectively.

[0086] y = LN(x + sublayer(x))

[0087] Here, LN —— Normalization layer; sublayer —— Multi-head self-attention or fully connected feed-forward network.

[0088]

[0089] Here, is the output of the multi-head attention module, is the output of the feed-forward neural network.

[0090] The decoder of KAformer replaces the simple MLP with Kolmogorov-Arnold Networks (KAN). KAN is created based on the Kolmogorov-Arnold theorem, and the specific mathematical formula is as follows:

[0091]

[0092] Here, φ p,m is a continuous univariate function. n, p, and q represent the number of network nodes, the number of top-ranked operators, and the number of underlying operators, respectively, which are controlled by the same variable N in the model.

[0093] The core layer of KAN is the KAN linear layer (KANlinear), which is represented by a B-spline function. The specific expression is:

[0094]

[0095] Here, B k (x) is the spline function, usually called the B-spline function.

[0096] The same channels are set at the beginning and end between the model and the sub-layer to solve the coordination problem of the input and output interfaces, and the KAformer model is built. Specifically: the output channels of the pre-LSTM network fully connected layer and the input channels of KAformer are both set to 11, and the output channels of the KAformer encoder and the input features of KANlinear in the decoder are both set to 10. The specific structural flow chart of the model is as Figure 3 shown.

[0097] 5) The specific process of the threshold setting method is as follows:

[0098] After the model constructed in step 4) is trained, the test set data is input into the model to obtain the predicted values of the target parameters, and the KLD calculation is performed with the true values. The calculation formula is:

[0099]

[0100] Here, P(x) represents the probability distribution of the predicted values, Q(x) represents the probability distribution of the true values, and KLD(P||Q) represents the KLD from P(x) to Q(x).

[0101] The D(P||Q) of the data sample points forms a detection curve characterizing the operating state of the wind turbine. Take part of the data in the training set, set a sliding time window with a window of 24 hours and a step size of 1 minute for sampling, and use the kernel probability density estimation method to calculate two safety threshold lines. The specific formula is as follows:

[0102]

[0103] Here, p(h) represents the density estimate value at point h, h = {h 1 , h 2 , …, h N}, h k is the k-th sample data point, N is the number of samples included in the set sliding time window, σ is the kernel function bandwidth coefficient, and K(·) is the kernel function. Here, the Gaussian kernel function is used, and its expression is as follows:

[0104]

[0105] Here, T 1 is the calculated warning threshold, 98% is the safety confidence level, T 2 is the alarm threshold, which takes the maximum value of p(h) in part of the training set data.

[0106] 6) The specific process of abnormal state detection is as follows:

[0107] According to the two threshold lines and the detection curve obtained in step 4), determine the evaluation indexes for the warning state and the alarm state. In Figure 4 , the horizontal black dotted line represents the warning threshold, the horizontal red solid line represents the alarm threshold, the vertical yellow solid line represents the position of the last sample point in the training set, the vertical black dotted line is the position of the sample point corresponding to the wind turbine fault shutdown and maintenance date. The warning state is the part where the detection curve outside the training set data (after the yellow line) crosses the warning threshold but does not cross the alarm threshold. The evaluation indexes include the number of warnings, warning sample points, warning duration and other indexes (among them, the fewer the number of warnings and the shorter the warning duration, the lower the misjudgment rate of the model); the alarm state is the part where the detection curve outside the training set data crosses the alarm threshold, and the alarm sample points, alarm duration and the first alarm date and other indexes are determined. If the alarm duration exceeds 60 minutes, reaching the severe fault level of the wind turbine, it is considered that the model can identify the abnormal state of the wind turbine. The indexes are shown in the following table, and the state detection diagram of the model is as Figure 4 shown. It can be seen that the LSTM-KAformer model has fewer pre-alarm times and shorter pre-alarm times, and can also give an alarm about 40 days in advance, with an alarm duration of 768 minutes, reaching the severe fault level of the wind turbine. In summary, the method designed by the present invention can well identify the abnormal operating state of the wind turbine with gearbox bearing faults.

[0108]

[0109] The present invention has the following advantages:

[0110] 1) The present invention designs a KAformer model. After replacing the encoder of the original ViT with a simple MLP by KAN, it can deeply extract the temporal and spatial features of SCADA data and accurately predict the target parameters.

[0111] 2) The present invention introduces KLD in mathematical statistics as an index to reflect the difference degree between the predicted value and the true value of the model, and combines KDE to set a threshold, improving the reliability and accuracy of state monitoring.

[0112] 3) The abnormal state detection method of wind turbines proposed by the present invention is verified by the SCADA data of wind turbines in an example wind farm, and can effectively detect the abnormal operating state of wind turbines.

Claims

1. A method for detecting abnormal state of a wind turbine generator system based on a KAformer model, characterized in that: The method comprises the following steps: a. Data preprocessing: clean and normalize the raw data of wind turbine operation obtained from the SCADA system, select target parameters and input parameters, and divide them into training sets and test sets; b. Construction of the detection model: Based on the original ViT, its encoder is replaced by KAN from a simple MLP to build a KAformer model, and a front LSTM network is added to the front of the KAformer model to obtain the required detection model; c. Model training: input the training set data into the detection model to train the detection model; d. Model testing: Input the test set data into the trained detection model to obtain the predicted value of the target parameter; perform KLD calculation on the predicted value and the true value of the target parameter to obtain the state monitoring curve, and then use the KDE method to calculate the early warning threshold and the alarm threshold; construct the abnormal state evaluation index based on the state monitoring curve, early warning threshold and alarm threshold, and judge the time period of abnormal operation of the wind turbine according to the evaluation index, so as to test the model; e. Use the qualified test model to detect the abnormal state of the wind turbine: input the operating data of the wind turbine to the qualified test model to obtain the state monitoring curve, and judge the time period of abnormal operation of the wind turbine according to the evaluation index.

2. The method for detecting abnormal state of a wind turbine generator system based on the KAformer model according to claim 1 is characterized in that: The specific process of data preprocessing is as follows: a1. Data cleaning and normalization: The specific steps for data cleaning using the quartile method are as follows: Arrange the values ​​of the curve into four intervals from small to large according to power and wind speed. The power of each wind speed interval is distributed within a quartile range from small to large, and the interquartile range (IQR, the difference between the third quartile Q3 and the first quartile Q1) is calculated, and then the upper limit value W is calculated. u ′ and the lower limit value W l ′, the formula is as follows: IQR=Q3-Q1 Above the upper limit W u ' or lower limit value W l The data of ′ are judged as outliers and are removed; SCADA data includes parameter data of wind speed, temperature, power, voltage and current, etc. The cleaned data needs to be normalized to convert various variables into the range of [0, 1]. The specific formula is as follows: a2. Selection of target parameters and input parameters: The target parameters are selected as parameters corresponding to the actual fault, and the input parameters are selected as parameters whose Spearman correlation coefficient with the target parameters is greater than 0.5; a3. Division of training set and test set: The data of the wind turbine set in the normal operating period is divided into the training set, and the data of the entire monitoring interval is divided into the test set.

3. The method for detecting abnormal state of a wind turbine generator system based on the KAformer model according to claim 1 is characterized in that: During the construction of the detection model, the construction method of each sub-layer is as follows: b1. Pre-LSTM layer: Set the number of pre-LSTM layers to 1 and the number of neurons to 64; b2.KAformer model: The KAformer model includes an embedding layer, an encoder, and a decoder. The embedding layer consists of a convolutional layer and a fully connected layer, and the input channel is set to 11; The encoder is composed of L encoding layers stacked in sequence, each encoding layer is composed of a self-attention mechanism layer, a residual connection and normalization layer, and a feedforward network layer, the MLP dimension is set to 512, the number of hidden layers is 256, and the multi-head attention mechanism is 8 heads; the decoder is composed of KANlinear, the number of nodes N is set to 10, and the spline parameter is 3.

4. The method for detecting abnormal state of a wind turbine generator system based on the KAformer model according to claim 1 is characterized in that: The training of the model consisted of 200 epochs, with a batch size of 256, a learning rate of 0.0001, a dropout of 0.1, an activation function of RELU, and an optimizer of Adam.

5. The method for detecting abnormal state of a wind turbine generator system based on the KAformer model according to claim 1 is characterized in that: The calculation method of the early warning threshold and the alarm threshold is as follows: Input the test set data into the trained detection model to obtain the predicted value of the target parameter, and perform KLD calculation on it and the true value: Where P(x) represents the probability distribution of the predicted value, Q(x) represents the probability distribution of the true value, KLD(P||Q) represents the KLD from P(x) to Q(x), and the KLD of the data sample points constitutes the detection curve that characterizes the operating status of the wind turbine. Take part of the training set data, set a sliding time window with a window of 24 hours and a step length of 1 hour for sampling, and use the KDE method to calculate the warning threshold and alarm threshold. The specific formula is: Where p(h) represents the density estimate at point h, h k is the kth sample data point, N is the number of samples contained in the set sliding time window, σ is the kernel function bandwidth coefficient, K(·) is the kernel function, and the security confidence s is set c , then the warning threshold T1 is calculated by the following formula: The alarm threshold T2 takes the maximum value of the training set data part p(h).

6. The abnormal state detection method of a wind turbine generator system based on the KAformer model according to claim 1 is characterized in that: The safety confidence level s c It is 98%.

7. The method for detecting abnormal state of a wind turbine generator system based on the KAformer model according to claim 1 is characterized in that: The kernel function K(·) is a Gaussian kernel function, and its expression is as follows:

8. The method for detecting abnormal state of a wind turbine generator system based on the KAformer model according to claim 1 is characterized in that: The evaluation indicators for determining the abnormal state include: the number of warnings, warning sample points, warning duration, alarm sample points, alarm duration and the first alarm date. The part of the detection curve that crosses the warning threshold but does not cross the alarm threshold is regarded as a warning state. The duration of the monitoring curve exceeding the alarm threshold for more than 60 minutes is regarded as an abnormal operating state of the wind turbine. The date corresponding to the sample point that first crosses the alarm threshold is the first alarm date.

Images