Wind power plant fan fault prediction method based on deep learning algorithm

Abstract

The invention discloses a wind power plant fan fault prediction method based on a deep learning algorithm, and the method comprises the following steps: obtaining and inputting fan operation state data, and obtaining a matrix capable of representing the characteristics of the operation state data; reducing the data size scale of a high-frequency signal sequence generated by the fan sensor by using a box line downsampling method; aiming at the fault sequence data of the positive sample, combining prior knowledge and observation data by using a Bayesian multivariate autoregression model; for negative sample data, normal behavior sequence data is mapped to approximate posterior probability distribution of potential variables output by an encoder by using a variational auto-encoder model in combination with an attention mechanism, a decoder samples from the distribution through a re-parameterization technique, and reconstructed data similar to input data distribution is generated; a self-adaptive threshold adjustment strategy is used, and a classification decision can be dynamically optimized by automatically adjusting the threshold of positive and negative reconstruction errors. According to the method, the fan fault prediction precision can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of wind turbine fault prediction, and in particular to a method for predicting wind turbine faults in a wind farm based on a deep learning algorithm. Background Art

[0002] The complexity of the mechanical structure of wind turbines and the uncertainty of the operating environment bring difficulties to the daily operation and maintenance services of large wind farms. Traditional physical model-driven fault diagnosis methods usually rely on accurate physical modeling and dynamic numerical simulation of mechanical structures. They are effective under ideal conditions, but it is difficult to achieve the expected prediction performance in actual scenarios. Machine learning algorithms can effectively characterize the deep characteristics of state data and are widely used in the diagnosis and prediction of wind turbine faults. However, in applications, the algorithm performance is degraded due to the loss of input sequence signals and the disappearance of gradients caused by the superposition of network layers. In particular, in the massive state time series data sets collected during the operation of wind turbines, the outliers when faults occur are relatively rare. That is, the extremely obvious imbalance of the data set in the fault prediction task brings great difficulties to traditional machine learning algorithms. How to effectively learn the respective characteristics of the data in the normal and abnormal states of the wind turbine at the same time will greatly improve the efficiency of learning and the accuracy of fault prediction, but it is also a major challenge. Summary of the invention

[0003] In view of the problems existing in the prior art, the object of the present invention is to provide a method for predicting wind turbine faults in a wind farm based on a deep learning algorithm.

[0004] To achieve the above purpose, the present invention uses the data distribution related statistics used in the box plot in each sliding time window based on the box line sampling method to deconstruct all the original features to reduce the sample size, and uses the Bayesian multivariate autoregression model and the attention variational autoencoder model to asymmetrically reconstruct the positive and negative samples respectively, combined with the adaptive threshold tree classification model, to effectively improve the fan fault prediction performance on the unbalanced data set. The adopted technical solution includes the following steps: Step 1: Obtain and input the fan operating status data, and obtain data that can characterize the fan operating status characteristics through a sliding window. , , and are the wind turbine state feature dimension and sliding window length respectively.

[0005] Step 2: Use the box sampling method to reduce the data size of the high-frequency signal sequence generated by the fan sensor: Step 2.1 The wind turbine operating status characteristic data in each sliding window Based on this, the statistics of data distribution characteristics are analyzed by combining the box plot to obtain the minimum value , lower quartile , median , upper quartile , maximum value and interquartile range ; Step 2.2 Deconstruct each feature of the sequence in each window into feature group data , specifically expressed as .

[0006] Step 3: Use the Bayesian multivariate autoregressive model to combine prior knowledge and observed data to reconstruct the positive sample fault sequence data with insufficient data volume (relative to negative samples, the amount of positive sample data is relatively small): Step 3.1 Assume that the positive sample sequence data of the fan The fan operating status characteristic data There is a linear correlation between them, and the error term is described by normal distribution. , then the positive sample sequence data is calculated through the Bayesian multivariate autoregressive model: ; in , is the regression coefficient, is the identity matrix, is the noise variance; Step 3.2 In the regression coefficient and noise variance Given the situation, we get the positive sample sequence data Conditional probability distribution with multivariate normal distribution ; Assumptions Follow normal prior distribution ; Then its posterior distribution is ; Step 3.3 Maximum a posteriori estimation MAP method obtains parameter estimates by minimizing the negative logarithmic posterior ; in and are the mean and covariance of the prior distribution, respectively.

[0007] Step 4: Use the variational autoencoder model combined with the attention mechanism to map the large and complex negative sample sequence data (relative to positive samples, the amount of negative sample data is relatively large) to the approximate posterior probability distribution of the latent vector output by the encoder. The decoder samples from this distribution using the reparameterization technique and generates reconstructed data similar to the input data distribution: Step 4.1 Assume that the latent variable follows a normal distribution Prior distribution and the generation probability Calculate representative observation data Marginal likelihood of total probability : ; By approximating the posterior distribution Instead of the true posterior distribution , according to the reconstruction error term that reflects the model's ability to reconstruct data in the latent space and a measure of the difference between the approximate posterior distribution and the prior distribution Divergence To maximize the approximate log marginal likelihood : ; Step 4.2: Use transformation matrix , and Input fan operating status characteristic data After linear transformation, we get the query matrix , key matrix Sum Matrix , and then calculate and The dot product of is used to obtain the similarity between each position and scale the similarity. All similarities are normalized into probability distribution through Softmax operation, which is applied to the value matrix to complete the weighted summation and generate the attention output. The multi-head attention mechanism repeats the attention mechanism Second, according to the output linear transformation matrix Concatenate the results to get the output of the multi-head attention mechanism ; ; in is the dimension of the key matrix; Step 4.3 will output the results The encoder adjusts the dimensions through multiple linear feedforward network layers to generate the distribution parameters of the latent space. and ; ; in is the activation function, which transforms the original linear transformation into a nonlinear transformation. and is the parameter matrix of the encoder linear feedforward network layer, and is the bias term; Step 4.4: Use a differentiable linear function to transform the noise that follows the normal distribution Combined with the network parameters to adjust the latent variables To take a sample: ; in represents element-wise multiplication; The decoder adjusts the sampled latent variables through multiple linear feedforward network layers to regenerate the observed data. ; ; in is the parameter matrix of the decoder linear feedforward network layer, is the bias term; Step 4.5: To maintain the smoothness and coherence of the model in the latent space, the divergence loss Introduced into the loss function middle: .

[0008] Step 5: Use the adaptive threshold adjustment strategy to set the thresholds for the positive and negative reconstruction errors and automatically adjust them to dynamically optimize the classification decision and output the fault detection results in a timely manner: Step 5.1 After completing the positive and negative sample reconstruction model training, freeze the model training parameters, and calculate the positive and negative deviations of the positive and negative training sets input into the two reconstruction models respectively; use the positive reconstruction error and negative reconstruction error The total error is measured by a linear weighted combination of : ; in and are the weight coefficients of positive and negative reconstruction errors respectively; Step 5.2 The fault label matrix composed of the output features corresponding to the fault labels is used to train a fault decision tree classification model; the decision tree model recursively divides the data set into smaller subsets according to the size of the information gain in fault classification, and gradually builds a tree structure: ; in is the information entropy, Features Information gain of Step 5.3 Output the classification results.

[0009] The beneficial effects of the present invention are as follows: machine learning, especially deep learning algorithms, have advantages in predicting faults of fans with complex structures, unstable operating environments and extremely unbalanced state data; the original high-frequency time series state data of the fan is compressed by the box-line undersampling method, while retaining the centralization trend and dispersion trend of the data, the data scale is reduced, providing support for effective training of the deep learning algorithm; by combining the Bayesian multivariate autoregressive model with the attention variational autoencoder model, the feature differences of positive and negative samples in fan faults can be effectively captured, thereby improving the fault prediction capability on unbalanced data sets; the Bayesian multivariate autoregressive model learns the distribution characteristics of small sample data by updating the posterior with the prior, while the attention variational autoencoder model solves the long dependency problem of time series data through the attention mechanism; the adaptive threshold adjustment strategy can dynamically optimize classification decisions, thereby achieving accurate fault detection in extremely unbalanced data sets. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0011] The present invention is further described below in conjunction with the accompanying drawings, but the protection scope of the present invention is not limited thereto.

[0012] like Figure 1 As shown, a method for predicting wind turbine failure in a wind farm based on a deep learning algorithm comprises the following steps: Step 1: Obtain and input the fan operating status data, and obtain data that can characterize the fan operating status characteristics through a sliding window. , , and are the wind turbine state feature dimension and sliding window length respectively.

[0013] The original wind turbine data set contains 34 feature dimensions, including time and identification information (timestamp), wind environment parameters (average wind speed within 10 seconds), blade and pitch control characteristics (average pitch angle within 1 second), power and energy characteristics (average active power within 10 seconds), mechanical and electrical characteristics (such as wind rotor speed, gearbox temperature, generator instantaneous current). The time node interval of the original data set is about 10 seconds, and the sliding time window length is set to 40 time steps. The initial data set of wind turbine status is shown in Table 1; Table 1 Input initial data set of wind turbine status

[0014] Step 2: Use the box sampling method to reduce the data size of the high-frequency signal sequence generated by the fan sensor: Step 2.1 The wind turbine operating status characteristic data in each sliding window Based on this, the statistics of data distribution characteristics are analyzed by combining the box plot to obtain the minimum value , lower quartile , median , upper quartile , maximum value and interquartile range ; Step 2.2 Deconstruct each feature of the sequence in each window into feature group data , specifically expressed as ; The deconstructed feature data are shown in Table 2; Table 2 Deconstructed feature data

[0015] Step 3: Use the Bayesian multivariate autoregressive model to combine prior knowledge and observed data to reconstruct the fault sequence data of insufficient positive samples: Step 3.1 Assume that the positive sample sequence data of the fan The fan operating status characteristic data There is a linear correlation between them, and the error term is described by normal distribution. , then the positive sample sequence data is calculated through the Bayesian multivariate autoregressive model: ; in , is the regression coefficient, is the identity matrix, is the noise variance.

[0016] The positive sample sequence data calculated by the Bayesian multivariate autoregression model is shown in Table 3; Table 3 Positive sample sequence data obtained by regression calculation

[0017] Step 3.2 In the regression coefficient and noise variance Given the situation, we get the positive sample sequence data Conditional probability distribution with multivariate normal distribution ; Assumptions Follow normal prior distribution ; Then its posterior distribution is ; Step 3.3 Maximum a posteriori estimation MAP method obtains parameter estimates by minimizing the negative logarithmic posterior ; in and are the mean and covariance of the prior distribution, respectively.

[0018] The parameter means and covariance values ​​obtained by minimizing the negative logarithmic posterior are shown in Tables 4 and 5; Table 4 Posterior means of parameters

[0019] Table 5 Posterior covariance values ​​of each parameter

[0020] Step 4: Use the variational autoencoder model combined with the attention mechanism to map the large and complex negative sample sequence data to the approximate posterior probability distribution of the latent vector output by the encoder. The decoder samples from this distribution using the reparameterization technique and generates reconstructed data similar to the input data distribution: Step 4.1 Assume that the latent variable follows a normal distribution Prior distribution and the generation probability Calculate representative observation data Marginal likelihood of total probability : ; By approximating the posterior distribution Instead of the true posterior distribution , according to the reconstruction error term that reflects the model's ability to reconstruct data in the latent space and a measure of the difference between the approximate posterior distribution and the prior distribution Divergence To maximize the approximate log marginal likelihood : ; Step 4.2: Use transformation matrix , and Input fan operating status characteristic data After linear transformation, we get the query matrix , key matrix Sum Matrix , and then calculate and The dot product of is used to obtain the similarity between each position and scale the similarity. All similarities are normalized into probability distribution through Softmax operation, which is applied to the value matrix to complete the weighted summation and generate the attention output. The multi-head attention mechanism repeats the attention mechanism Second, according to the output linear transformation matrix Concatenate the results to get the output of the multi-head attention mechanism ; ; in is the dimension of the key matrix; The attention output results are shown in Table 6; Table 6 Attention output results

[0021] Step 4.3 will output the results The encoder adjusts the dimensions through multiple linear feedforward network layers to generate the distribution parameters of the latent space. and ; ; in is the activation function, which transforms the original linear transformation into a nonlinear transformation. and is the parameter matrix of the encoder linear feedforward network layer, and is the bias term; Step 4.4: Use a differentiable linear function to transform the noise that follows the normal distribution Combined with the network parameters to adjust the latent variables To take a sample: ; in represents element-wise multiplication; The decoder adjusts the sampled latent variables through multiple linear feedforward network layers to regenerate the observed data. ; ; in is the parameter matrix of the decoder linear feedforward network layer, is the bias term; The regenerated observation data are shown in Table 7; Table 7 Regenerated observation data

[0022] Step 4.5: To maintain the smoothness and coherence of the model in the latent space, the divergence loss Introduced into the loss function middle: .

[0023] Step 5: Use the adaptive threshold adjustment strategy to set the thresholds for the positive and negative reconstruction errors and automatically adjust them to dynamically optimize the classification decision and output the fault detection results in a timely manner: Step 5.1 After completing the positive and negative sample reconstruction model training, freeze the model training parameters, and calculate the positive and negative deviations of the positive and negative training sets input into the two reconstruction models respectively; use the positive reconstruction error and negative reconstruction error The total error is measured by a linear weighted combination of : ; in and are the weight coefficients of positive and negative reconstruction errors respectively; Step 5.2 The fault label matrix composed of the output features corresponding to the fault labels is used to train a fault decision tree classification model; the decision tree model recursively divides the data set into smaller subsets according to the size of the information gain in fault classification, and gradually builds a tree structure: ; in is the information entropy, Features Information gain of Step 5.3 Output the classification results.

[0024] The classification results output in a certain period are shown in Table 8; Table 8 Output classification results

[0025] The accuracy, recall, precision and F1 score are selected as the classification evaluation indicators. The wind turbine status dataset of the same period is used for training, verification and testing. The comparison with other mainstream prediction methods is shown in Table 9: Table 9 Comparison of prediction results

[0026] As can be seen from Table 9, the proposed method performs well in processing unbalanced data sets, and has significant improvements in accuracy, recall, precision, and F1 score compared to traditional supervised learning models (such as Transformer, LSTM), autoencoder models (such as Autoencoder), and related hybrid models.

Claims

1. A method for predicting wind turbine failure in a wind farm based on a deep learning algorithm, characterized in that: The following steps are involved: Step 1: Obtain and input the wind turbine operating status data to obtain a matrix that can characterize the characteristics of the operating status data; Step 2: Use the box-line downsampling method to reduce the data volume of the high-frequency signal sequence generated by the fan sensor; Step 3: Use the Bayesian multivariate autoregressive model to combine prior knowledge and observed data to reconstruct the positive sample fault sequence data with relatively small data volume; Step 4: Use the variational autoencoder model combined with the attention mechanism to map the relatively large amount of negative sample sequence data to the latent vector approximation posterior probability distribution output by the encoder. The decoder uses the reparameterization technique to sample from the posterior probability distribution and generate reconstructed data similar to the input data distribution. Step 5: Use the adaptive threshold adjustment strategy to set the thresholds for the positive and negative reconstruction errors and automatically adjust them to dynamically optimize the classification decision and output the fault detection results in a timely manner.

2. A method for predicting wind turbine failure in a wind farm based on a deep learning algorithm as claimed in claim 1, characterized in that The specific operation process of step 1 includes the following steps: after obtaining and inputting the fan operation status data, the data that can characterize the fan operation status characteristics is obtained through the sliding window. , , and are the wind turbine state feature dimension and sliding window length respectively.

3. A method for predicting wind turbine failure in a wind farm based on a deep learning algorithm as claimed in claim 2, characterized in that The specific operation process of step 2 includes the following steps: Step 2.1: The wind turbine operating status characteristic data in each sliding window Based on this, the statistics of data distribution characteristics are analyzed by combining the box plot to obtain the minimum value , lower quartile , median , upper quartile , maximum value and interquartile range ; Step 2.2: Deconstruct each feature of the sequence in each window into feature group data , specifically expressed as 。 4. A method for predicting wind turbine failure in a wind farm based on a deep learning algorithm as claimed in claim 3, characterized in that The specific operation process of step 3 includes the following steps: Step 3.1, assuming that the positive sample sequence data of the fan Characteristics of fan operation status There is a linear correlation between them, and the error term is described by normal distribution. , then the positive sample sequence data is calculated through the Bayesian multivariate autoregressive model: ； in , is the regression coefficient, is the identity matrix, is the noise variance; Step 3.2, in the regression coefficient and noise variance Given the situation, we get the positive sample sequence data Conditional probability distribution with multivariate normal distribution ； Assumptions Follow normal prior distribution ； Then its posterior distribution is ； Step 3.3, the maximum a posteriori estimation MAP method obtains parameter estimates by minimizing the negative logarithmic posterior ； in and are the mean and covariance of the prior distribution, respectively.

5. A method for predicting wind turbine failure in a wind farm based on a deep learning algorithm as claimed in claim 4, characterized in that The specific operation process of step 4 includes the following steps: Step 4.1, assume that the latent variable follows a normal distribution Prior distribution and the generation probability Calculate representative observation data Marginal likelihood of total probability : ； By approximating the posterior distribution Instead of the true posterior distribution , according to the reconstruction error term that reflects the model's ability to reconstruct data in the latent space and a measure of the difference between the approximate posterior distribution and the prior distribution Divergence To maximize the approximate log marginal likelihood : ； Step 4.2, respectively through the transformation matrix , and Input fan operating status characteristic data After linear transformation, we get the query matrix , key matrix Sum Matrix , and then calculate and The dot product of is used to obtain the similarity between each position and scale the similarity. All similarities are normalized into probability distribution through Softmax operation, which is applied to the value matrix to complete the weighted summation and generate the attention output. Multi-head attention mechanism repeats through the attention mechanism Second, according to the output linear transformation matrix Concatenate the results to get the output of the multi-head attention mechanism ; ； in is the dimension of the key matrix; Step 4.3, the output result will be The encoder adjusts the dimensions through multiple linear feedforward network layers to generate the distribution parameters of the latent space. and ; ； in is the activation function, which transforms the original linear transformation into a nonlinear transformation. and is the parameter matrix of the encoder linear feedforward network layer, and is the bias term; Step 4.4: Use a differentiable linear function to transform the noise that follows the normal distribution Combined with the network parameters to adjust the latent variables To sample: ； in represents element-wise multiplication; The decoder adjusts the sampled latent variables through multiple linear feedforward network layers to regenerate the observed data. ; ； in is the parameter matrix of the decoder linear feedforward network layer, is the bias term; Step 4.5: To maintain the smoothness and coherence of the model in the latent space, the divergence loss Introduced into the loss function middle: 。 6. A method for predicting wind turbine failure in a wind farm based on a deep learning algorithm as claimed in claim 5, characterized in that The specific operation process of step 5 includes the following steps: Step 5.1, freeze the model training parameters after completing the positive and negative sample reconstruction model training, and calculate the positive and negative deviations of the positive and negative training sets input into the two reconstruction models respectively; use the positive reconstruction error and negative reconstruction error The total error is measured by a linear weighted combination of : ； in and are the weight coefficients of positive and negative reconstruction errors respectively; Step 5.2, the fault label matrix composed of the output features corresponding to the fault labels is used to train a fault decision tree classification model; the decision tree model recursively divides the data set into smaller subsets according to the size of the information gain in fault classification, and gradually builds a tree structure: ； in is the information entropy, Features Information gain of Step 5.3, output the classification results.

Images