A method for fault detection of a wind turbine gearbox

Abstract

The present invention discloses a method for fault detection of a wind turbine gearbox, which includes two stages: offline model training and online fault detection. The offline model training stage includes operating state classification and classification-regression model training. The operating state classification uses an unsupervised classification method for operating states based on dynamic time warping (DTW), and the classification-regression model training uses a classification-regression algorithm combined with a sampling weight allocation strategy. The online fault detection stage uses a fault detection confidence calculation strategy based on a residual overrun percentage warning mechanism. During the detection process of the present invention, the fault detection method is data-driven and can fully consider different operating states of the gearbox; through independent detection of different operating states, fault detection under complex operating conditions of the gearbox is achieved. In addition, by using the sampling weight allocation strategy, the influence brought by the data imbalance problem can be overcome, making the classification model perform better and achieving a better online fault detection effect.

Description

Technical Field

[0001] The present invention belongs to the technical field of fault monitoring, and particularly relates to a method for fault detection of a wind turbine gearbox. Background Art

[0002] In a wind turbine, the gearbox is a key subsystem connecting the blades and the generator. Faults in the gearbox will directly affect the operating states of other key components in the unit, causing great economic losses and even leading to safety accidents. Therefore, the condition monitoring and fault detection of the gearbox are crucial for improving the operating reliability of the wind turbine and reducing the operation and maintenance costs.

[0003] The existing methods for fault detection of wind turbine gearboxes are mainly divided into model-based or data-driven methods. The model-based method aims to establish a physical model of the main components in the gearbox to describe the evolution process of faults. However, due to the complex structure of the gearbox, the existence of non-linear relationships between variables, and the uncertainty of external disturbances, it is difficult to accurately describe the true state of the system with the established mathematical model, which hinders the application of the model in actual scenarios. In contrast, the data-driven method can describe the operating state of the wind turbine based on the data measured by sensors without relying on an accurate mathematical model, so it has become the mainstream method recommended in relevant literature. In addition, the supervisory control and data acquisition (SCADA) system has been widely applied to wind turbines, providing a reliable data basis for the data-driven method.

[0004] However, there are still some challenges and problems in the data-driven method. The complex and changeable wind conditions lead to the changeable operating states of wind turbines, especially bringing obvious complexity to the operating state of the gearbox. The data-driven model needs to grasp these different states. In addition, since different wind conditions have different proportions in historical data, it is necessary to study how to balance them to avoid unbalanced model training. Summary of the Invention

[0005] To make up for the deficiencies of the existing technology, the present invention provides a method for fault detection of a wind turbine gearbox to solve the problems that the complex and changeable wind conditions lead to the changeable operating states of wind turbines, especially bringing obvious complexity to the operating state of the gearbox, and the different proportions of different wind conditions in historical data lead to data imbalance, etc.

[0006] The described method for fault detection of a wind turbine gearbox includes two stages: offline model training and online fault detection. The offline model training stage includes operating state classification and classification-regression model training. The operating state classification uses an unsupervised classification method for operating states based on dynamic time warping (DTW), and the classification-regression model training uses a classification-regression algorithm combined with a sampling weight allocation strategy. The online fault detection stage uses a fault detection confidence calculation strategy based on a residual overrun percentage warning mechanism.

[0007] Further, the unsupervised classification method for operating states based on dynamic time warping (DTW) in the offline model training stage includes state variable selection, data sample regularization, DTW distance calculation, and K-means clustering.

[0008] Specifically, the specific steps of the unsupervised classification method for operating states based on dynamic time warping (DTW) are as follows:

[0009] Select the wind speed v, generator speed g, and output active power P as the operating state variables of the wind turbine, denoted as X o , and perform regularization on the operating state variables X of the wind turbine o :

[0010]

[0011] In the formula, is the mean vector, and σ is the regularization coefficient;

[0012] Select the state variables of the wind turbine during stable power generation under stable wind conditions Use a moving time window to segment the data to form a reference sample:

[0013]

[0014] In the formula, W is the length of the time window, and each sample needs to calculate the distance from this reference sample to compare the correlation;

[0015] Select the dynamic time warping (DTW) to calculate the distance between the sample and the reference sample, that is:

[0016]

[0017] In the formula, W is the warping path, K is the length of the warping path, and each state variable in the sample is calculated to obtain the corresponding DTW distance. For the i-th sample, the finally obtained distance is expressed as:

[0018]

[0019] In the formula, and They respectively represent the DTW distances corresponding to the wind speed, generator speed, and output active power in the sample.

[0020] The K-means clustering algorithm is used to cluster the DTW distances, and the decrease value of the sum of squared errors SSE of the clustering results is statistically calculated as the number of clusters increases. The minimum number of clusters corresponding to when SSE no longer significantly decreases is taken as the number of clusters of K-means.

[0021] Furthermore, the classification-regression algorithm combining the sampling weight allocation strategy in the offline model training stage includes training sample sampling weight allocation, XGBoost classification model, GRU regression model, and residual threshold setting.

[0022] Specifically, the specific steps of the classification-regression algorithm of the sampling weight allocation strategy are as follows:

[0023] For the DTW distance of the sample, the extreme gradient boosting XGBoost decision tree model is adopted and trained, the probability that the input sample falls into each category is calculated, and the sampling weight allocation strategy is adopted. The sampling weight α of the sample in the i-th cluster i is calculated as follows:

[0024]

[0025] In the formula, n o is the number of categories obtained by clustering, and N i is the number of samples in the i-th category;

[0026] After training the XGBoost classification model, for each category corresponding to the training sample classification result, a gated recurrent unit GRU is respectively selected to construct a regression model for training. The bearing temperature is selected as the output variable y of the GRU regression model, and the relevant variables X selected through prior knowledge are used as the input of the model. Therefore, n o GRU regression models are respectively trained for the categories separated by clustering, and a total of n o models are trained; for each model, the residual between the predicted value and the true value of the bearing temperature is respectively calculated, and the 3-sigma criterion is selected to determine the threshold of the residual under the i-th category:

[0027] U i = μ i + 3σ i

[0028] In the formula, μ i and σ i are respectively the average value and standard deviation of the residual sequence in the i-th category.

[0029] Furthermore, the fault detection confidence calculation strategy based on the residual overrun percentage warning mechanism in the online fault detection process stage includes the regularization of the data to be detected, the identification of the XGBoost operating state + GRU bearing temperature prediction, the residual overrun warning value, and the fault detection confidence.

[0030] Specifically, the specific steps of the fault detection confidence calculation strategy based on the residual overrun percentage warning mechanism are as follows:

[0031] For a certain real-time data sample X r , three variables, namely the wind speed v, the generator speed g, and the output active power P, are selected and regularized as the input of the XGBoost classification model. The output of the model is the probability that the sample belongs to each category:

[0032]

[0033] In the formula, P i is the probability that X r belongs to the i-th category;

[0034] Traverse the real-time data sample X in chronological order r The residual sequence P of the bearing temperature d , when P d % of the residuals exceed the limit, the warning value is set to 1. If this ratio is still not reached after all traversals, the warning value is 0; the value of P d can be specified according to the actual situation, so the GRU regression model for each category will give the warning value for that category:

[0035]

[0036] X r The fault detection confidence D of r The calculation formula is as follows:

[0037]

[0038] That is, this confidence D r is expressed as the sum of the probabilities of each warning category.

[0039] Compared with the prior art, the present invention has the following advantages: The present invention includes two stages of offline model training and online fault detection. It adopts an unsupervised classification method for operating states based on dynamic time warping, a classification - regression algorithm combined with a sampling weight allocation strategy, and a fault detection scoring strategy based on a residual over - limit percentage warning mechanism. During the detection process, the fault detection method is data - driven and can fully consider different operating states of the gearbox. Through independent detection of different operating states, fault detection under complex operating conditions of the gearbox is achieved. In addition, by adopting a sampling weight allocation strategy, opposite sampling weights are assigned according to the number of samples in each operating state category. The training sample set obtained through the weighted mechanism sampling can be evenly distributed, overcoming the influence brought by the data imbalance problem, making the classification model perform better and achieving a better online fault detection effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 is a block diagram of the steps of the method of the present invention;

[0041] Figure 2 is an example diagram of the sampling weight allocation strategy of the method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0042] To enable those skilled in the art to more clearly understand the technical solutions of the present application, the following further describes a method for fault detection of a wind turbine gearbox according to the present invention with reference to the accompanying drawings.

[0043] As Figure 1 shown, a method for fault detection of a wind turbine gearbox includes two stages of offline model training and online fault detection. The offline model training stage includes operating state classification and classification - regression model training. The operating state classification adopts an unsupervised classification method for operating states based on dynamic time warping (DTW), and the classification - regression model training adopts a classification - regression algorithm combined with a sampling weight allocation strategy. The online fault detection stage adopts a fault detection confidence calculation strategy based on a residual over - limit percentage warning mechanism. The specific details are as follows:

[0044] (1) The unsupervised classification method for operating states based on dynamic time warping in the offline model training stage includes state variable selection, data sample regularization, DTW distance calculation, and K - means clustering, which are specifically as follows:

[0045] Select the wind speed v, generator speed g, and output active power P as the operating state variables of the wind turbine, denoted as X o , and regularize the operating state variables X o of the wind turbine:

[0046]

[0047] In the formula, is the mean vector, and σ is the regularization coefficient.

[0048] To measure the correlation between different samples, the state variables of the wind turbine during stable power generation under stable wind conditions are selected The data is segmented using a moving time window to form a reference sample:

[0049]

[0050] In the formula, W is the length of the time window, and each sample needs to calculate the distance from this reference sample to compare the correlation. Since this correlation is independent of the phase difference of the samples, the Euclidean distance cannot be used to measure it. In this application, dynamic time warping (DTW) is selected to calculate the distance between the sample and the reference sample, that is:

[0051]

[0052] In the formula, W is the warping path, K is the length of the warping path, and the DTW distance corresponding to each state variable in the sample is calculated separately. For the i-th sample, the finally obtained distance is expressed as:

[0053]

[0054] In the formula, and respectively represent the DTW distances corresponding to the wind speed, generator speed, and output active power in this sample.

[0055] Since samples in different operating states will show different DTW distances, in this application, the K-means clustering algorithm is used to cluster the DTW distances to achieve the effect of classifying different operating states. To determine the number of clusters, it is necessary to gradually increase the number of clusters during the offline model training stage, and count the decrease in the sum of squared errors (SSE) of the clustering results as the number of clusters increases. The minimum number of clusters corresponding to when the SSE no longer decreases significantly is taken as the number of clusters of K-means.

[0056] (2) The classification-regression algorithm combined with the sampling weight allocation strategy in the offline model training stage includes training sample sampling weight allocation, XGBoost classification model, GRU regression model, and residual threshold setting, as follows:

[0057] After classifying the operating states through an unsupervised classification method for operating states based on dynamic time warping, for the DTW distances of the samples, an extreme gradient boosting (XGBoost) decision tree model is adopted and trained to calculate the probability that the input samples fall into each category. To overcome the problem of unbalanced data samples in different categories, a sampling weight allocation strategy is adopted, and the sampling weight α of the samples in the i-th cluster is i calculated as follows:

[0058]

[0059] where n o is the number of categories obtained by clustering, and N i is the number of samples in the i-th category. Through this sampling weight allocation strategy, the number of training samples collected from each category reaches balance (as shown in the appendix Figure 2 ), ensuring the learning and classification effects of the XGBoost model.

[0060] After training the XGBoost classification model, for each category corresponding to the classification results of the training samples, a gated recurrent unit (GRU) is respectively selected to construct a regression model for training. Since the bearing temperature of the gearbox reflects the state of the gearbox and many potential faults are manifested as abnormal increases in the bearing temperature, the bearing temperature is selected as the output variable y of the GRU regression model, and other relevant variables X selected through prior knowledge are used as the inputs of the model. Therefore, a GRU regression model is trained for each of the n o categories separated by clustering, and a total of n o models are trained. For each model, the residual between the predicted value and the true value of the bearing temperature is calculated respectively. To characterize the change range of the residual under the normal working state of the gearbox, the present invention selects the 3-sigma criterion to determine the threshold of the residual under the i-th category:

[0061] U i = μ i + 3σ i

[0062] where μ i and σ i are respectively the mean and standard deviation of the residual sequence in the i-th category.

[0063] It can be seen that by adopting the sampling weight allocation strategy, opposite sampling weights are assigned according to the number of samples in each operating state category. The training sample set obtained through the weighted mechanism sampling can be evenly distributed, overcoming the influence brought by the problem of data imbalance, making the classification model perform better and achieving a better online fault detection effect.

[0064] (3) After completing the offline model training process, enter the online detection process. The fault detection confidence calculation strategy based on the residual overrun percentage warning mechanism adopted in the online fault detection process stage includes regularization of the data to be detected, XGBoost running state recognition + GRU bearing temperature prediction, residual overrun warning value, and fault detection confidence.

[0065] For a certain real-time data sample X r , select 3 variables: wind speed v, generator speed g, and output active power P, and after regularization, use them as the input of the XGBoost classification model. The output of the model is the probability that the sample belongs to each category:

[0066]

[0067] In the formula, P i is the probability that X r belongs to the i-th category.

[0068] The GRU regression model for each category will respectively give the predicted value of the bearing temperature based on X r . When comparing the prediction residual with the previously determined residual threshold U i , considering the volatility of the variables and avoiding the influence of outliers, the fault detection confidence calculation strategy based on the residual overrun percentage warning mechanism adopted in the present invention is specifically described as follows:

[0069] Traverse the residual sequence of the bearing temperature of X r in chronological order. When the residual of P d % exceeds the limit, set the warning value to 1; if this proportion is still not reached after all traversals, the warning value is 0. The value of P d can be specified according to the actual situation, so the GRU regression model for each category will give the warning value for this category:

[0070]

[0071] The fault detection confidence D r of X r is calculated as follows:

[0072]

[0073] This confidence level is expressed as the sum of the probabilities of each warning category, reflecting the confidence level of the occurrence of the fault, and can provide more detailed and accurate fault detection information compared to the binary classification result.

[0074] The fault detection method of the present invention is data-driven and can fully consider different operating states of the gearbox; through independent detection of different operating states, fault detection under complex operating conditions of the gearbox is achieved.

[0075] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting a wind turbine gearbox fault, characterized in that: It includes two stages: offline model training and online fault detection. The offline model training stage includes operation status classification and classification-regression model training. The operation status classification adopts the operation status unsupervised classification method based on dynamic time warping DTW, and the classification-regression model training adopts the classification-regression algorithm combined with the sampling weight allocation strategy. The online fault detection stage adopts the fault detection confidence calculation strategy based on the residual over-limit percentage warning mechanism. The unsupervised classification method of the running status based on dynamic time warping (DTW) in the offline model training stage includes state variable selection, data sample regularization, DTW distance calculation and K-means clustering; the classification-regression algorithm combined with the sampling weight allocation strategy in the offline model training stage includes training sample sampling weight allocation, XGBoost classification model, GRU regression model and residual threshold setting; The specific steps of the classification-regression algorithm of the sampling weight allocation strategy include: according to the DTW distance of the sample, the extreme gradient boosting XGBoost decision tree model is adopted and trained, the probability of the input sample falling into each category is calculated, and the sampling weight allocation strategy is adopted to allocate the sampling weight α of the sample in the i-th cluster. i The calculation is as follows: Where n o is the number of categories obtained by clustering, N i is the number of samples in the i-th category; The fault detection confidence calculation strategy based on the residual over-limit percentage warning mechanism in the online fault detection process stage includes regularization of the data to be detected, XGBoost operating status identification + GRU bearing temperature prediction, residual over-limit warning value and fault detection confidence.

2. A method for detecting a wind turbine gearbox fault according to claim 1, characterized in that: The specific steps of the unsupervised classification method of running status based on dynamic time warping DTW are as follows: Select wind speed v, generator speed g, and output active power P as the operating state variables of the wind turbine group, denoted as X o , the wind turbine operating state variable X o Regularization: In the formula, is the mean vector, σ is the regularization coefficient; Select the state variables of the wind turbine when it generates stable power under stable wind conditions Use a moving time window to segment the data and form a benchmark sample: Where W is the length of the time window, and each sample needs to be compared with the benchmark sample by calculating the distance; Dynamic time warping (DTW) is selected to calculate the distance between the sample and the benchmark sample, that is: Where W is the regular path, K is the length of the regular path, and each state variable in the sample is calculated to obtain the corresponding DTW distance. For the i-th sample, the final distance is expressed as: In the formula, and They represent the DTW distances corresponding to the wind speed, generator speed, and output active power in the sample respectively; The K-means clustering algorithm is used to cluster the DTW distance. The decrease value of the sum of square errors (SSE) of the clustering results as the number of clusters increases is counted. The minimum number of clusters corresponding to the SSE no longer decreases significantly is taken as the number of K-means clusters.

3. A method for detecting a wind turbine gearbox fault according to claim 2, characterized in that: The specific steps of the classification-regression algorithm of the sampling weight allocation strategy also include: after training the XGBoost classification model, for each category corresponding to the classification result of the training sample, a gated recurrent unit GRU is selected to build a regression model for training, the bearing temperature is selected as the output variable y of the GRU regression model, and the related variable X selected by prior knowledge is used as the input of the model. o A GRU regression model was trained for each category, and a total of n o models; for each model, the residual between the predicted value and the true value of the bearing temperature is calculated, and the 3-sigma criterion is selected to determine the threshold of the residual under the i-th category: U i =μ i +3s i In the formula, μ i and σ i are the mean and standard deviation of the residual sequence in the ith category, respectively.

4. A method for detecting a wind turbine gearbox fault according to claim 3, characterized in that: The specific steps of the fault detection confidence calculation strategy based on the residual over-limit percentage warning mechanism are as follows: For a real-time data sample X r , select three variables: wind speed v, generator speed g, and output active power P, and use them as the input of the XGBoost classification model after regularization. The model outputs the probability that the sample belongs to each category: Where P i Yes X r The probability of belonging to the i-th category; Traverse real-time data samples X in chronological order r The residual sequence P of the bearing temperature d , when P d When the residual of % exceeds the limit, the warning value is set to 1. If this ratio is still not reached after all traversals, the warning value is 0; d The value of can be specified according to the actual situation, so the GRU regression model of each category will give the warning value of that category: X r The fault detection confidence D r The calculation formula is as follows: That is, the confidence D r It is expressed as the sum of the probabilities of each warning category.

Images