Wind turbine condition monitoring system and method based on cointegration analysis

Abstract

Disclosed is a wind turbine state monitoring system and method based on cointegration analysis, which processes wind speed, impeller speed, generator speed and generator power obtained at multiple predetermined time points through a multi-scale neighborhood feature extraction module to obtain multi-scale wind speed feature vectors, multi-scale impeller speed feature vectors, multi-scale generator speed feature vectors and multi-scale power feature vectors, and calculates corresponding responsiveness estimates to obtain a first cointegration transfer matrix, a second cointegration transfer matrix and a third cointegration transfer matrix, which are then arranged into a three-dimensional input tensor, corrected through each feature matrix along the channel dimension of the convolutional neural network to obtain a corrected classification feature map, and finally classified by a classifier to obtain a classification result indicating whether the state of the wind turbine is normal. In this way, the correlation characteristics of external environmental factors and internal factors of offshore wind turbines are mined, and the detection accuracy and reliability are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of new energy, and more particularly to a wind turbine state monitoring system and method based on cointegration analysis. BACKGROUND

[0002] Wind turbine generators are developing towards high power and high efficiency, so unexpected accidents of the wind turbine generators will cause high repair costs and huge losses to the enterprise, and even threaten people's lives and property, so it is particularly important to identify and handle faults early.

[0003] At present, the realization of wind turbine state monitoring is to monitor mechanical parameters, and common monitoring objects include oil monitoring, temperature monitoring, vibration monitoring, etc. The supervisory control and data acquisition (SCADA) system is continuously introduced into the state monitoring of wind turbine generators. The SCADA system collects a large amount of data (such as speed, temperature, power and angle, etc.), so analyzing the key data can realize the state monitoring of the wind turbine generator. However, it is very difficult to analyze and interpret the large amount of SCADA data of the wind turbine generator, and the reliability of the data analysis result is insufficient due to the influence of the environment and the operation factors of the wind turbine generator.

[0004] Therefore, an optimized wind turbine state monitoring system is expected to monitor the abnormal state of the wind turbine generator based on the operation of the wind turbine generator and external factors. SUMMARY

[0005] In order to solve the above technical problems, the present application is proposed. The embodiments of the present application provide a wind turbine state monitoring system and method based on cointegration analysis, which processes the wind speed, impeller speed, generator speed and generator power at a plurality of predetermined time points to obtain a multi-scale wind speed feature vector, a multi-scale impeller speed feature vector, a multi-scale generator speed feature vector and a multi-scale power feature vector, and calculates the corresponding responsiveness estimates to obtain a first cointegration transfer matrix, a second cointegration transfer matrix and a third cointegration transfer matrix, and then arranges them into a three-dimensional input tensor, and corrects each feature matrix along the channel dimension after the convolutional neural network to obtain a corrected classification feature map, and then uses a classifier to obtain a classification result representing whether the state of the wind turbine is normal. In this way, the correlation characteristics of external environmental factors and internal factors of offshore wind turbines are mined, and the detection accuracy and reliability are improved.

[0006] According to an aspect of the present application, a wind turbine state monitoring system based on cointegration analysis is provided, which comprises:

[0007] a data acquisition and monitoring control module configured to acquire wind speed, impeller speed, generator speed and generator power at a plurality of predetermined time points;

[0008] a vectorization module configured to arrange the wind speed, impeller speed, generator speed and generator power at the plurality of predetermined time points into wind speed input vectors, impeller speed input vectors, generator speed input vectors and power input vectors, respectively;

[0009] a multi-scale time series feature extraction module configured to input the wind speed input vectors, the impeller speed input vectors, the generator speed input vectors and the power input vectors into a multi-scale neighborhood feature extraction module to obtain multi-scale wind speed feature vectors, multi-scale impeller speed feature vectors, multi-scale generator speed feature vectors and multi-scale power feature vectors, respectively;

[0010] a cointegration analysis module configured to calculate the responsiveness estimates of the multi-scale impeller speed feature vectors, the multi-scale generator speed feature vectors and the multi-scale power feature vectors with respect to the multi-scale wind speed feature vectors to obtain a first cointegration transition matrix, a second cointegration transition matrix and a third cointegration transition matrix, respectively;

[0011] a correlation feature extraction module configured to arrange the first cointegration transition matrix, the second cointegration transition matrix and the third cointegration transition matrix into a three-dimensional input tensor, and then use a convolutional neural network with convolution kernels transposed to each other in adjacent layers to obtain a classification feature map;

[0012] a correction module configured to correct each feature matrix of the classification feature map along the channel dimension based on a reference feature vector to obtain a corrected classification feature map, wherein the reference feature vector is a channel feature vector obtained by calculating the global mean of each feature matrix of the classification feature map along the channel dimension; and

[0013] a state monitoring result generation module configured to pass the corrected classification feature map through a classifier to obtain a classification result, wherein the classification result is used to indicate whether the state of the wind turbine is normal.

[0014] In the wind turbine state monitoring system based on cointegration analysis, the multi-scale time sequence feature extraction module comprises: a first convolution coding unit, configured to input the wind speed input vector, the impeller speed input vector, the generator speed input vector and the power input vector into a first convolution layer of the multi-scale neighborhood feature extraction module respectively to obtain a first neighborhood scale wind speed correlation feature vector, a first neighborhood scale impeller speed correlation feature vector, a first neighborhood scale generator speed correlation feature vector and a first neighborhood scale power correlation feature vector, wherein the first convolution layer has a first one-dimensional convolution kernel with a first length; a second convolution coding unit, configured to input the wind speed input vector, the impeller speed input vector, the generator speed input vector and the power input vector into a second convolution layer of the multi-scale neighborhood feature extraction module respectively to obtain a second neighborhood scale wind speed correlation feature vector, a second neighborhood scale impeller speed correlation feature vector, a second neighborhood scale generator speed correlation feature vector and a second neighborhood scale power correlation feature vector, wherein the second convolution layer has a second one-dimensional convolution kernel with a second length, and the first length is different from the second length; and a concatenation unit, configured to concatenate the first neighborhood scale wind speed correlation feature vector and the second neighborhood scale wind speed correlation feature vector to obtain the multi-scale wind speed feature vector, concatenate the first neighborhood scale impeller speed correlation feature vector and the second neighborhood scale impeller speed correlation feature vector to obtain the multi-scale impeller speed feature vector, concatenate the first neighborhood scale generator speed correlation feature vector and the second neighborhood scale generator speed correlation feature vector to obtain the multi-scale generator speed feature vector, and concatenate the first neighborhood scale power correlation feature vector and the second neighborhood scale power correlation feature vector to obtain the multi-scale power feature vector.

[0015] In the wind turbine state monitoring system based on cointegration analysis, the cointegration analysis module is further configured to: calculate the responsiveness estimates of the multi-scale impeller speed feature vector, the multi-scale generator speed feature vector and the multi-scale power feature vector with respect to the multi-scale wind speed feature vector respectively by using the following formula to obtain the first cointegration transfer matrix, the second cointegration transfer matrix and the third cointegration transfer matrix; wherein the formula is:

[0016] M i =M j *M1

[0017] wherein M irespectively represent the multi-scale impeller speed feature vector, the multi-scale generator speed feature vector and the multi-scale power feature vector, M1 represents the multi-scale wind speed feature vector, and M1, M2 and M3 respectively represent the first cointegration transfer matrix, the second cointegration transfer matrix and the third cointegration transfer matrix.

[0018] In the above wind turbine state monitoring system based on cointegration analysis, the correction module comprises: a reference feature vector generation unit configured to calculate the global mean of each feature matrix along the channel dimension of the classification feature map to obtain the reference feature vector; an optimization factor calculation unit configured to calculate the optimization factor corresponding to the feature value of each position in the reference feature vector based on the mean and variance of the feature value set of all positions in the reference feature vector; and a correction unit configured to weight the corresponding feature matrix in the classification feature map with the optimization factor corresponding to the feature value of each position in the reference feature vector as the weight to obtain the corrected classification feature map.

[0019] In the above wind turbine state monitoring system based on cointegration analysis, the optimization factor calculation unit is further configured to calculate the optimization factor corresponding to the feature value of each position in the reference feature vector based on the mean and variance of the feature value set of all positions in the reference feature vector according to the following formula: wherein the formula is:

[0020]

[0021] wherein v i represents the feature value of the i-th position of the reference feature vector, and μ and σ respectively represent the mean and variance of the feature value set of all positions in the reference feature vector, exp(·) represents the exponential operation of the feature value, and the exponential operation of the feature value represents the calculation of the natural exponential function value with the feature value as the power.

[0022] According to another aspect of the present application, a wind turbine state monitoring method based on cointegration analysis is also provided, which comprises:

[0023] obtaining wind speed, impeller speed value, generator speed and generator power at a plurality of predetermined time points;

[0024] arranging the wind speed, impeller speed value, generator speed and generator power at the plurality of predetermined time points into a wind speed input vector, an impeller speed input vector, a generator speed input vector and a power input vector, respectively;

[0025] inputting the wind speed input vector, the impeller speed input vector, the generator speed input vector and the power input vector into a multi-scale neighborhood feature extraction module respectively to obtain a multi-scale wind speed feature vector, a multi-scale impeller speed feature vector, a multi-scale generator speed feature vector and a multi-scale power feature vector;

[0026] calculating a responsiveness estimate of the multi-scale impeller speed feature vector, the multi-scale generator speed feature vector and the multi-scale power feature vector with respect to the multi-scale wind speed feature vector respectively to obtain a first cointegration transfer matrix, a second cointegration transfer matrix and a third cointegration transfer matrix;

[0027] arranging the first cointegration transfer matrix, the second cointegration transfer matrix and the third cointegration transfer matrix into a three-dimensional input tensor and then using a convolutional neural network with transposed convolution kernels in adjacent layers to obtain a classification feature map;

[0028] correcting each feature matrix along the channel dimension of the classification feature map based on a reference feature vector to obtain a corrected classification feature map, wherein the reference feature vector is a channel feature vector obtained by calculating the global mean of each feature matrix along the channel dimension of the classification feature map; and

[0029] passing the corrected classification feature map through a classifier to obtain a classification result, which is used to represent whether the state of the wind turbine is normal.

[0030] Compared with the prior art, the wind turbine state monitoring system and method based on cointegration analysis provided in the present application process the wind speed, impeller speed value, generator speed and generator power obtained at a plurality of predetermined time points through a multi-scale neighborhood feature extraction module to obtain a multi-scale wind speed feature vector, a multi-scale impeller speed feature vector, a multi-scale generator speed feature vector and a multi-scale power feature vector, and calculate corresponding responsiveness estimates to obtain a first cointegration transfer matrix, a second cointegration transfer matrix and a third cointegration transfer matrix, then arrange them into a three-dimensional input tensor and pass them through a convolutional neural network to correct each feature matrix along the channel dimension to obtain a corrected classification feature map, and then pass the corrected classification feature map through a classifier to obtain a classification result used to represent whether the state of the wind turbine is normal. In this way, the correlation characteristics of external environmental factors and internal factors of offshore wind turbines are mined, and the detection and judgment accuracy and reliability are improved. BRIEF DESCRIPTION OF DRAWINGS

[0031] The above and other objects, features and advantages of the present application will become more apparent from the following detailed description thereof taken in conjunction with the accompanying drawings, in which: The accompanying drawings provide exemplary embodiments of the application and serve to illustrate the principle of the present application. The embodiments described herein are fully to be covered within the scope of the present application, and provide further understanding of the present application together with the embodiments of the present application, and do not limit the present application. In the drawings, like reference numerals refer to like parts or steps throughout the several views.

[0032] Figure 1 FIG. 1 illustrates an application scenario diagram of a co-integration analysis based wind turbine condition monitoring system according to an embodiment of the present application.

[0033] Figure 2 FIG. 2 illustrates a block diagram of a co-integration analysis based wind turbine condition monitoring system according to an embodiment of the present application.

[0034] Figure 3 FIG. 3 illustrates a system architecture diagram of a co-integration analysis based wind turbine condition monitoring system according to an embodiment of the present application.

[0035] Figure 4 FIG. 4 illustrates a block diagram of a multi-scale time series feature extraction module in a co-integration analysis based wind turbine condition monitoring system according to an embodiment of the present application.

[0036] Figure 5 FIG. 5 illustrates a block diagram of a correlation feature extraction module in a co-integration analysis based wind turbine condition monitoring system according to an embodiment of the present application.

[0037] Figure 6 FIG. 6 illustrates a block diagram of a correction module in a co-integration analysis based wind turbine condition monitoring system according to an embodiment of the present application.

[0038] Figure 7 FIG. 7 illustrates a flowchart of a co-integration analysis based wind turbine condition monitoring method according to an embodiment of the present application. DETAILED DESCRIPTION

[0039] Hereinafter, example embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the example embodiments described herein.

[0040] SUMMARY

[0041] It can be understood that, due to the influence of the environment and its operation factors on the wind turbine during the monitoring of the state of the wind turbine, the reliability of the data analysis result is insufficient. Therefore, in order to improve the detection judgment accuracy and reliability during the state monitoring of the wind turbine, the implicit feature information of the external environmental elements and the internal factors of the offshore wind turbine, and the correlation characteristics of the external environmental elements and the internal factors of the offshore wind turbine are selected. That is, in the technical solution of the present application, the external environmental elements can be represented by wind speed data, and the internal factors of the offshore wind turbine can be represented by impeller speed value, generator speed and generator power data, and then the respective multi-scale neighborhood correlation feature representation is extracted by using a deep neural network model, and the data co-integration of the high-dimensional feature space is performed by using the responsiveness estimation between the data features, thereby improving the effectiveness and reliability of the abnormal state monitoring of the wind turbine.

[0042] Specifically, in the technical solution of the present application, first, the wind speed, impeller speed value, generator speed and generator power of a plurality of predetermined time points are obtained by each sensor. Then, in order to obtain the distribution information of the entire data in the time dimension, so as to facilitate the subsequent deep correlation feature mining, the wind speed, impeller speed value, generator speed and generator power of the plurality of predetermined time points are respectively arranged into a wind speed input vector, an impeller speed input vector, a generator speed input vector and a power input vector.

[0043] It should be appreciated that the convolutional neural network is originally a model applied in the field of image, but the idea of local feature extraction can also be applied to time series data analysis. For example, a time series convolution structure with a convolution kernel size of 3, for wind speed time series data input, the convolution kernel moves in the form of a sliding window along the time dimension, and outputs the weighted sum of the data in each wind speed time series segment. Each convolution unit stacks multiple convolution kernels to output multi-dimensional wind speed features. Because the wind speed will show different patterns in different time periods, such as the wind speed will be strong for a while and weak for a while in different time periods. The feature extraction of the field is to mine the deep correlation features of the data in different time spans. A large convolution kernel extracts features from a large-scale time series neighborhood of the wind speed, where each wind speed value in the neighborhood has a smaller impact, thereby reducing the fluctuations of the wind speed input data and reducing the impact of noise points on the output features of the wind speed. However, a large-scale convolution kernel reduces the difference in numerical changes, which can easily lead to over-smoothing, causing the output features of the wind speed to lose discriminability. On the contrary, a small-scale convolution kernel can better preserve the information in the wind speed input data, but it is also more susceptible to interference from noise points. Therefore, considering the characteristics of convolution of different scales, different sizes of convolution units are combined to extract features of different wind speed time series scales. Then, feature fusion is completed by feature splicing to obtain multi-scale wind speed neighborhood features. It is worth mentioning that similarly, for the impeller speed, the generator speed and the power input data, multi-scale neighborhood correlation features can also be extracted based on this method.

[0044] That is, specifically, in the technical solution of the present application, a multi-scale neighborhood feature extraction module with one-dimensional convolution kernels of different scales is further used to perform one-dimensional convolution coding on the wind speed input vector, the impeller speed input vector, the generator speed input vector and the power input vector, respectively, and then the feature vectors corresponding to the two one-dimensional convolution kernels of different scales are concatenated to obtain the multi-scale wind speed feature vector, the multi-scale impeller speed feature vector, the multi-scale generator speed feature vector and the multi-scale power feature vector. In particular, in this way, the multi-scale neighborhood correlation feature representation of the dynamic implicit features of the wind speed, the impeller speed, the generator speed and the power is extracted, respectively, so that the output features contain both smoothed features and original input features, avoiding information loss and improving the accuracy of subsequent classification. In particular, in other specific examples of the present application, the multi-scale neighborhood feature extraction module can also include a larger number of one-dimensional convolution layers using one-dimensional convolution kernels of different lengths to extract correlation features in different scales of the neighborhood, which is not limited by the present application.

[0045] Since the wind speed data is an external environmental factor, the impeller speed, the generator speed and the power are internal factors of the offshore wind turbine, and the feature scales between the multiscale neighborhood correlation features of the wind speed, the impeller speed, the generator speed and the power are not the same, while the multiscale neighborhood correlation features of the impeller speed, the generator speed and the power can be respectively regarded as responsive features to the multiscale neighborhood correlation features of the wind speed in the high-dimensional feature space. Therefore, in order to better fuse these feature information, further calculate the responsiveness estimation of the multiscale impeller speed feature vector, the multiscale generator speed feature vector and the multiscale power feature vector to the multiscale wind speed feature vector respectively to carry out data co-integration of the high-dimensional feature space to obtain a first co-integration transfer matrix, a second co-integration transfer matrix and a third co-integration transfer matrix.

[0046] Then, after arranging the first co-integration transfer matrix, the second co-integration transfer matrix and the third co-integration transfer matrix into a three-dimensional input tensor to integrate the data co-integration feature information of respective data in the high-dimensional feature space, a convolutional neural network with convolution kernels of adjacent layers being transposed to each other is used to process the three-dimensional input tensor to obtain a classification feature map. In particular, here, the adjacent convolution layers of the neural network use convolution kernels that are transposed to each other, so that the update of the network parameters and the search of the network parameter structure suitable for the specific data structure can be updated at the same time during training, thereby improving the accuracy of subsequent classification.

[0047] In particular, in the technical solution of the present application, when the first co-integration transfer matrix, the second co-integration transfer matrix and the third co-integration transfer matrix are arranged into a three-dimensional input tensor to obtain the classification feature map, the convolution kernels of adjacent layers that are transposed to each other can extract specific structures within the matrix dimension, so it is expected that the classification feature map can have better expression in the channel dimension.

[0048] Based on this, first, the global mean of each feature matrix along the channel of the classification feature map is calculated to obtain a channel feature vector, and then a deep recursive squeeze-and-excitation optimization factor of each channel is calculated based on a reference feature vector, specifically:

[0049]

[0050] where v i represents the feature value of the i-th position of the reference feature vector, and μ and σ respectively represent the mean and variance of the set of feature values of all positions in the reference feature vector.

[0051] Here, the deep recursive squeeze-excitation optimization factor can activate the along-channel depth recursion of the feature distribution based on the statistical characteristics of the feature set along the channel, thereby inferring the depth distribution of the features at each channel sampling depth thereof, and adopting a channel squeeze-excitation mechanism composed of a ReLU-Sigmoid function to obtain a depth confidence value with channel attention enhancement, which is used as a coefficient to weight the feature matrix of each channel of the classification feature map, which can improve the expression confidence of the high-dimensional feature flow pattern in the depth direction, thereby improving the classification accuracy. In this way, the abnormal state of the wind turbine can be monitored to avoid the occurrence of unexpected accidents of the wind turbine set, thereby ensuring the normal operation of the wind turbine.

[0052] Based on this, the application provides a wind turbine state monitoring system based on cointegration analysis, which comprises: a data acquisition and monitoring control module, used to obtain wind speed, impeller speed value, generator speed and generator power at a plurality of predetermined time points; a vectorization module, used to arrange the wind speed, impeller speed value, generator speed and generator power at the plurality of predetermined time points into a wind speed input vector, an impeller speed input vector, a generator speed input vector and a power input vector respectively; a multi-scale time sequence feature extraction module, used to input the wind speed input vector, the impeller speed input vector, the generator speed input vector and the power input vector into a multi-scale neighborhood feature extraction module respectively to obtain a multi-scale wind speed feature vector, a multi-scale impeller speed feature vector, a multi-scale generator speed feature vector and a multi-scale power feature vector; a cointegration analysis module, used to calculate the responsiveness estimation of the multi-scale impeller speed feature vector, the multi-scale generator speed feature vector and the multi-scale power feature vector relative to the multi-scale wind speed feature vector respectively to obtain a first cointegration transfer matrix, a second cointegration transfer matrix and a third cointegration transfer matrix; a correlation feature extraction module, used to arrange the first cointegration transfer matrix, the second cointegration transfer matrix and the third cointegration transfer matrix into a three-dimensional input tensor, and then use a convolutional neural network with transposed convolution kernels in adjacent layers to obtain a classification feature map; a correction module, used to correct each feature matrix along the channel dimension of the classification feature map based on a reference feature vector to obtain a corrected classification feature map, wherein the reference feature vector is a channel feature vector obtained by calculating the global mean of each feature matrix along the channel dimension of the classification feature map; and a state monitoring result generation module, used to pass the corrected classification feature map through a classifier to obtain a classification result, wherein the classification result is used to indicate whether the state of the wind turbine is normal.

[0053] Figure 1 The application scenario diagram of the wind turbine state monitoring system based on cointegration analysis is shown in FIG. 1. Figure 1As shown, in this application scenario, firstly, wind speed, impeller speed value, generator speed and generator power at multiple predetermined time points are acquired through various sensors (for example, Se1, Se2, Se3, Se4 as shown in Figure 1 FIG. 2). Figure 1 Then, the wind speed, impeller speed value, generator speed and generator power at the multiple predetermined time points are input into a server (for example, S as shown in FIG. 3) in which a wind turbine state monitoring algorithm based on cointegration analysis is deployed, wherein the server is capable of processing the input information with the wind turbine state monitoring algorithm based on cointegration analysis to obtain a classification result indicating whether the state of the wind turbine is normal.

[0054]

[0055] Exemplary System

[0056] Figure 2 FIG. 1 illustrates a block diagram of a wind turbine state monitoring system based on cointegration analysis according to an embodiment of the present application. As shown in FIG. 1, the wind turbine state monitoring system based on cointegration analysis according to an embodiment of the present application comprises a wind turbine 1, a server 2 and a user terminal 3. Figure 2As shown, the wind turbine state monitoring system 100 based on cointegration analysis according to the embodiment of the present application comprises: a data acquisition and monitoring control module 110, configured to acquire wind speed, impeller speed value, generator speed and generator power at a plurality of predetermined time points; a vectorization module 120, configured to arrange the wind speed, the impeller speed value, the generator speed and the generator power at the plurality of predetermined time points into a wind speed input vector, an impeller speed input vector, a generator speed input vector and a power input vector respectively; a multi-scale time sequence feature extraction module 130, configured to input the wind speed input vector, the impeller speed input vector, the generator speed input vector and the power input vector into a multi-scale neighborhood feature extraction module respectively to obtain a multi-scale wind speed feature vector, a multi-scale impeller speed feature vector, a multi-scale generator speed feature vector and a multi-scale power feature vector; a cointegration analysis module 140, configured to calculate the responsiveness estimation of the multi-scale impeller speed feature vector, the multi-scale generator speed feature vector and the multi-scale power feature vector with respect to the multi-scale wind speed feature vector respectively to obtain a first cointegration transfer matrix, a second cointegration transfer matrix and a third cointegration transfer matrix; a correlation feature extraction module 150, configured to arrange the first cointegration transfer matrix, the second cointegration transfer matrix and the third cointegration transfer matrix into a three-dimensional input tensor, and then use a convolutional neural network with transposed convolution kernels in adjacent layers to obtain a classification feature map; a correction module 160, configured to correct each feature matrix along the channel dimension of the classification feature map based on a reference feature vector to obtain a corrected classification feature map, wherein the reference feature vector is a channel feature vector obtained by calculating the global mean of each feature matrix along the channel dimension of the classification feature map; and a state monitoring result generation module 170, configured to pass the corrected classification feature map through a classifier to obtain a classification result, wherein the classification result is used to indicate whether the state of the wind turbine is normal.

[0057] Figure 3 FIG. 1 illustrates a system architecture diagram of the wind turbine state monitoring system 100 based on cointegration analysis according to the embodiment of the present application. As shown in FIG. 1, the wind turbine state monitoring system 100 based on cointegration analysis according to the embodiment of the present application comprises: a data acquisition and monitoring control module 110, configured to acquire wind speed, impeller speed value, generator speed and generator power at a plurality of predetermined time points; a vectorization module 120, configured to arrange the wind speed, the impeller speed value, the generator speed and the generator power at the plurality of predetermined time points into a wind speed input vector, an impeller speed input vector, a generator speed input vector and a power input vector respectively; a multi-scale time sequence feature extraction module 130, configured to input the wind speed input vector, the impeller speed input vector, the generator speed input vector and the power input vector into a multi-scale neighborhood feature extraction module respectively to obtain a multi-scale wind speed feature vector, a multi-scale impeller speed feature vector, a multi-scale generator speed feature vector and a multi-scale power feature vector; a cointegration analysis module 140, configured to calculate the responsiveness estimation of the multi-scale impeller speed feature vector, the multi-scale generator speed feature vector and the multi-scale power feature vector with respect to the multi-scale wind speed feature vector respectively to obtain a first cointegration transfer matrix, a second cointegration transfer matrix and a third cointegration transfer matrix; a correlation feature extraction module 150, configured to arrange the first cointegration transfer matrix, the second cointegration transfer matrix and the third cointegration transfer matrix into a three-dimensional input tensor, and then use a convolutional neural network with transposed convolution kernels in adjacent layers to obtain a classification feature map; a correction module 160, configured to correct each feature matrix along the channel dimension of the classification feature map based on a reference feature vector to obtain a corrected classification feature map, wherein the reference feature vector is a channel feature vector obtained by calculating the global mean of each feature matrix along the channel dimension of the classification feature map; and a state monitoring result generation module 170, configured to pass the corrected classification feature map through a classifier to obtain a classification result, wherein the classification result is used to indicate whether the state of the wind turbine is normal. Figure 3As shown, in the system architecture of the wind turbine state monitoring system 100 based on cointegration analysis, firstly, wind speed, impeller speed value, generator speed and generator power at multiple predetermined time points are obtained, and are arranged into wind speed input vector, impeller speed input vector, generator speed input vector and power input vector respectively. Then, multi-scale wind speed feature vector, multi-scale impeller speed feature vector, multi-scale generator speed feature vector and multi-scale power feature vector are obtained by using multi-scale neighborhood feature extraction module, and the responsiveness estimation of the multi-scale impeller speed feature vector, the multi-scale generator speed feature vector and the multi-scale power feature vector with respect to the multi-scale wind speed feature vector is calculated respectively to obtain a first cointegration transfer matrix, a second cointegration transfer matrix and a third cointegration transfer matrix. Then, the first cointegration transfer matrix, the second cointegration transfer matrix and the third cointegration transfer matrix are arranged into a three-dimensional input tensor, and then a convolutional neural network using convolution kernels transposed to each other is used to obtain a classification feature map, and each feature matrix along the channel dimension of the classification feature map is corrected to obtain a corrected classification feature map. Then, the corrected classification feature map is passed through a classifier to obtain a classification result, which is used to represent whether the state of the wind turbine is normal.

[0058] In the above-mentioned wind turbine state monitoring system 100 based on cointegration analysis, the data acquisition and monitoring control module 110 is used to obtain wind speed, impeller speed value, generator speed and generator power at multiple predetermined time points. It should be understood that, when monitoring the state of the wind turbine, the wind turbine will be affected by the environment and its operating factors, which will make the data analysis result unreliable. Therefore, in order to improve the detection judgment accuracy and reliability when monitoring the state of the wind turbine, the hidden feature information of external environmental elements and offshore wind turbine internal factors, as well as the correlation characteristics of the external environmental elements and the offshore wind turbine internal factors are selected. That is, in the technical solution of the present application, the external environmental elements can be represented by wind speed data, and the offshore wind turbine internal factors can be represented by impeller speed value, generator speed and generator power data.

[0059] In the wind turbine state monitoring system 100 based on the co-integration analysis, the vectorization module 120 is configured to arrange the wind speed, the impeller speed value, the generator speed and the generator power at the plurality of predetermined time points into a wind speed input vector, an impeller speed input vector, a generator speed input vector and a power input vector respectively. In order to obtain the distribution information of the entire data in the time dimension, the wind speed, the impeller speed value, the generator speed and the generator power at the plurality of predetermined time points are further arranged into a wind speed input vector, an impeller speed input vector, a generator speed input vector and a power input vector respectively.

[0060] In the wind turbine state monitoring system 100 based on the co-integration analysis, the multi-scale time sequence feature extraction module 130 is configured to input the wind speed input vector, the impeller speed input vector, the generator speed input vector and the power input vector into a multi-scale neighborhood feature extraction module to obtain a multi-scale wind speed feature vector, a multi-scale impeller speed feature vector, a multi-scale generator speed feature vector and a multi-scale power feature vector. It should be understood that the convolutional neural network is originally a model applied in the image field, but the idea of local feature extraction can also be applied to time sequence data analysis. For example, a time sequence convolution structure with a convolution kernel size of 3, for wind speed time sequence data input, the convolution kernel moves in the form of a sliding window along the time dimension, and outputs the weighted sum of the data in each wind speed time sequence segment. Each convolution unit stacks multiple convolution kernels to output multi-dimensional wind speed features. Since the wind speed will present different patterns in different time periods, for example, the wind speed will be strong for a while and weak for a while in different time periods. The neighborhood feature extraction is to perform deep correlation feature mining on the data in different time spans, and a large convolution kernel extracts features from a large-scale time sequence neighborhood of the wind speed, in which the influence of each wind speed value in the neighborhood is smaller, thereby reducing the fluctuations of the wind speed input data and reducing the influence of noise points on the output features of the wind speed. However, the large-scale convolution kernel reduces the difference in numerical changes, which easily leads to the problem of over-smoothing, causing the output features of the wind speed to lose discriminability. On the contrary, a small-scale convolution kernel can better preserve the information in the wind speed input data, but it is also more susceptible to interference from noise points. Therefore, considering the characteristics of different scales of convolution, different sizes of convolution units are combined to extract features of different wind speed time scales. Then, feature fusion is completed by feature splicing, thereby obtaining multi-scale wind speed neighborhood features. It is worth mentioning that, similarly, the multi-scale neighborhood correlation features of the impeller speed, the generator speed and the power input data can also be extracted based on this method.

[0061] That is, specifically, in the technical solution of the present application, further using the convolution layer of the multi-scale neighborhood feature extraction module with one-dimensional convolution kernels of different scales respectively encodes the wind speed input vector, the impeller speed input vector, the generator speed input vector and the power input vector, and then concatenates the feature vectors corresponding to the two different scales of one-dimensional convolution kernels to obtain the multi-scale wind speed feature vector, the multi-scale impeller speed feature vector, the multi-scale generator speed feature vector and the multi-scale power feature vector. In particular, in this way, the multi-scale neighborhood correlation feature representation of the dynamic implicit features of the wind speed, the impeller speed, the generator speed and the power is extracted respectively, so that the output features contain both smoothed features and original input features, avoiding information loss and improving the accuracy of subsequent classification. In particular, in other specific examples of the present application, the multi-scale neighborhood feature extraction module can also include a larger number of one-dimensional convolution layers using one-dimensional convolution kernels of different lengths for different scale neighborhood correlation feature extraction, which is not limited by the present application.

[0062] Figure 4 Fig. 1 illustrates a block diagram of a multi-scale time series feature extraction module in a cointegration analysis-based wind turbine state monitoring system according to an embodiment of the present application. As shown in the figure, the multi-scale time series feature extraction module includes a multi-scale neighborhood feature extraction module and a multi-scale feature concatenation module. Figure 4As shown, the multi-scale time sequence feature extraction module 130 comprises: a first convolutional coding unit 131, configured to input the wind speed input vector, the impeller speed input vector, the generator speed input vector and the power input vector into a first convolutional layer of the multi-scale neighborhood feature extraction module respectively to obtain a first neighborhood scale wind speed correlation feature vector, a first neighborhood scale impeller speed correlation feature vector, a first neighborhood scale generator speed correlation feature vector and a first neighborhood scale power correlation feature vector, wherein the first convolutional layer has a first one-dimensional convolution kernel with a first length; a second convolutional coding unit 132, configured to input the wind speed input vector, the impeller speed input vector, the generator speed input vector and the power input vector into a second convolutional layer of the multi-scale neighborhood feature extraction module respectively to obtain a second neighborhood scale wind speed correlation feature vector, a second neighborhood scale impeller speed correlation feature vector, a second neighborhood scale generator speed correlation feature vector and a second neighborhood scale power correlation feature vector, wherein the second convolutional layer has a second one-dimensional convolution kernel with a second length, and the first length is different from the second length; and a concatenation unit 133, configured to concatenate the first neighborhood scale wind speed correlation feature vector and the second neighborhood scale wind speed correlation feature vector to obtain the multi-scale wind speed feature vector, concatenate the first neighborhood scale impeller speed correlation feature vector and the second neighborhood scale impeller speed correlation feature vector to obtain the multi-scale impeller speed feature vector, concatenate the first neighborhood scale generator speed correlation feature vector and the second neighborhood scale generator speed correlation feature vector to obtain the multi-scale generator speed feature vector, and concatenate the first neighborhood scale power correlation feature vector and the second neighborhood scale power correlation feature vector to obtain the multi-scale power feature vector.

[0063] In one example, in the above-mentioned wind turbine state monitoring system 100 based on cointegration analysis, the first convolutional coding unit 131 is further configured to use the first convolutional layer of the multi-scale neighborhood feature extraction module to perform one-dimensional convolutional coding on the wind speed input vector, the impeller speed input vector, the generator speed input vector and the power input vector respectively according to the following formula to obtain the first neighborhood scale wind speed correlation feature vector, the first neighborhood scale impeller speed correlation feature vector, the first neighborhood scale generator speed correlation feature vector and the first neighborhood scale power correlation feature vector.

[0064] The formula is:

[0065]

[0066] wherein a is the width of the first convolution kernel in the x direction, F(a) is the first convolution kernel parameter vector, G(x-a) is the local vector matrix operated with the convolution kernel function, w is the size of the first convolution kernel, X represents the wind speed input vector, the impeller speed input vector, the generator speed input vector and the power input vector.

[0067] In one example, in the above-mentioned wind turbine state monitoring system 100 based on cointegration analysis, the second convolutional coding unit 132 is further configured to use the second convolutional layer of the multi-scale neighborhood feature extraction module to perform one-dimensional convolutional coding on the wind speed input vector, the impeller speed input vector, the generator speed input vector and the power input vector respectively according to the following formula to obtain the second neighborhood scale wind speed correlation feature vector, the second neighborhood scale impeller speed correlation feature vector, the second neighborhood scale generator speed correlation feature vector and the second neighborhood scale power correlation feature vector.

[0068] wherein the formula is:

[0069]

[0070] wherein b is the width of the second convolution kernel in the x direction, F(b) is the second convolution kernel parameter vector, G(x-b) is the local vector matrix operated with the convolution kernel function, m is the size of the second convolution kernel, X represents the wind speed input vector, the impeller speed input vector, the generator speed input vector and the power input vector.

[0071] In the aforementioned cointegration analysis-based wind turbine condition monitoring system 100, the cointegration analysis module 140 is configured to respectively calculate the responsiveness estimates of the multi-scale impeller speed eigenvector, the multi-scale generator speed eigenvector, and the multi-scale power eigenvector relative to the multi-scale wind speed eigenvector to obtain a first cointegration transfer matrix, a second cointegration transfer matrix, and a third cointegration transfer matrix. Since the wind speed data is an external environmental factor, the impeller speed, the generator speed, and the power are internal factors of the offshore wind turbine, and the characteristic scales of the multi-scale neighborhood correlation features of the wind speed, the impeller speed, the generator speed, and the power are different, the multi-scale neighborhood correlation features of the impeller speed, the generator speed, and the power can be viewed as responsiveness features to the multi-scale neighborhood correlation features of the wind speed in a high-dimensional feature space. Therefore, in order to better integrate these characteristic information, the responsiveness estimates of the multi-scale impeller speed eigenvector, the multi-scale generator speed eigenvector and the multi-scale power eigenvector relative to the multi-scale wind speed eigenvector are further calculated to perform data cointegration in the high-dimensional feature space, so as to obtain the first cointegration transfer matrix, the second cointegration transfer matrix and the third cointegration transfer matrix.

[0072] In one example, in the above-mentioned wind turbine condition monitoring system 100 based on cointegration analysis, the cointegration analysis module 140 is further used to: calculate the responsiveness estimates of the multi-scale impeller speed eigenvector, the multi-scale generator speed eigenvector, and the multi-scale power eigenvector relative to the multi-scale wind speed eigenvector using the following formulas to obtain the first cointegration transfer matrix, the second cointegration transfer matrix, and the third cointegration transfer matrix;

[0073] Wherein, the formula is:

[0074] M i =M j *M1

[0075] Among them, M i They represent the multi-scale impeller speed eigenvector, the multi-scale generator speed eigenvector and the multi-scale power eigenvector respectively, M1 represents the multi-scale wind speed eigenvector, and M represents the first cointegration transfer matrix, the second cointegration transfer matrix and the third cointegration transfer matrix respectively.

[0076] In the aforementioned wind turbine condition monitoring system 100 based on cointegration analysis, the correlation feature extraction module 150 is configured to arrange the first, second, and third cointegration transfer matrices into a three-dimensional input tensor, and then process the resulting data using a convolutional neural network with mutually transposed convolution kernels in adjacent layers to obtain a classification feature map. Specifically, after arranging the first, second, and third cointegration transfer matrices into a three-dimensional input tensor to integrate the data cointegration feature information of their respective data in a high-dimensional feature space, the three-dimensional input tensor is processed using a convolutional neural network with mutually transposed convolution kernels in adjacent layers to obtain a classification feature map. Specifically, adjacent convolutional layers of the neural network use mutually transposed convolution kernels, thereby enabling simultaneous updating of network parameters and searching for a network parameter structure suitable for a specific data structure during training, thereby improving the accuracy of subsequent classification.

[0077] Figure 5 FIG2 is a block diagram of a correlation feature extraction module in a wind turbine condition monitoring system based on cointegration analysis according to an embodiment of the present application. Figure 5 As shown, the associated feature extraction module 150 includes: a shallow feature map extraction unit 151, which is used to extract a shallow feature map from the Mth layer of the convolutional neural network using mutually transposed convolution kernels in the adjacent layers, where M is an even number; a deep feature map extraction unit 152, which is used to extract a deep feature map from the Nth layer of the convolutional neural network using mutually transposed convolution kernels in the adjacent layers, where N is an even number and is greater than 2 times of M; and a feature map fusion unit 153, which is used to fuse the shallow feature map and the deep feature map to generate the classification feature map.

[0078] In the above-mentioned wind turbine condition monitoring system 100 based on cointegration analysis, the correction module 160 is used to correct each feature matrix along the channel dimension of the classification feature map based on a reference feature vector to obtain a corrected classification feature map, wherein the reference feature vector is a channel feature vector obtained by calculating the global mean of each feature matrix along the channel dimension of the classification feature map. In particular, in the technical solution of the present application, when the first cointegration transfer matrix, the second cointegration transfer matrix, and the third cointegration transfer matrix are arranged as a three-dimensional input tensor to obtain the classification feature map, the convolution kernels of adjacent layers that are mutually transposed can extract specific structures within the matrix dimension, so it is expected that the classification feature map can have better expression in the channel dimension.

[0079] Based on this, first, the global mean of each feature matrix along the channel of the classification feature map is calculated to obtain a channel feature vector, and then a deep recursive squeeze-and-excitation optimization factor of each channel is calculated based on the reference feature vector. Here, the deep recursive squeeze-and-excitation optimization factor can activate the channel depth recursion of the feature distribution based on the statistical characteristics of the feature set along the channel, thereby inferring the depth distribution of the features at each channel sampling depth thereof, and adopting a channel squeeze-and-excitation mechanism composed of a ReLU-Sigmoid function to obtain a deep confidence value enhanced by channel attention, so as to weight the feature matrix of each channel of the classification feature map with the value as a coefficient, which can improve the expression confidence of the deep direction of the high-dimensional feature flow type of the classification feature map, thereby improving the accuracy of classification. In this way, the abnormal state of the wind turbine can be monitored to avoid the occurrence of unexpected accidents of the wind turbine set, thereby ensuring the normal operation of the wind turbine.

[0080] Figure 6 Fig. 1 illustrates a block diagram of a correction module in a cointegration analysis-based wind turbine state monitoring system according to an embodiment of the present application. As shown in the figure, the correction module 160 comprises a reference feature vector generation unit 161 configured to calculate the global mean of each feature matrix along the channel dimension of the classification feature map to obtain the reference feature vector; an optimization factor calculation unit 162 configured to calculate the optimization factor corresponding to the feature value at each position in the reference feature vector based on the mean and variance of the feature value set at all positions in the reference feature vector; and a correction unit 163 configured to weight the corresponding feature matrix in the classification feature map with the optimization factor corresponding to the feature value at each position in the reference feature vector as a weight to obtain the corrected classification feature map. Figure 6

[0081] In one example, in the cointegration analysis-based wind turbine state monitoring system 100 described above, the optimization factor calculation unit 162 is further configured to calculate the optimization factor corresponding to the feature value at each position in the reference feature vector based on the mean and variance of the feature value set at all positions in the reference feature vector according to the following formula:

[0082] wherein the formula is:

[0083]

[0084] wherein v i represents the feature value at the i-th position of the reference feature vector, and μ and σ represent the mean and variance of the feature value set at all positions in the reference feature vector, respectively, and exp(·) represents the exponential operation of the feature value, which represents the calculation of the natural exponential function value with the feature value as the power. ​

[0085] In the above-mentioned wind turbine condition monitoring system 100 based on cointegration analysis, the condition monitoring result generating module 170 is used to pass the corrected classification feature map through a classifier to obtain a classification result, and the classification result is used to indicate whether the condition of the wind turbine is normal.

[0086] In one example, in the above-mentioned wind turbine condition monitoring system 100 based on cointegration analysis, the condition monitoring result generating module 170 is further configured to: the classifier processes the corrected classification feature map using the following formula to generate a classification result, wherein the formula is: softmax{(W n ,B n ):…:(W1,B1)|Project(F)}, where Project(F) represents projecting the corrected classification feature map into a vector, W1 to W n is the weight matrix of each fully connected layer, B1 to B n Represents the bias matrix of each fully connected layer.

[0087] In summary, the wind turbine condition monitoring system 100 based on cointegration analysis according to the embodiment of the present application is explained. It processes the wind speed, impeller speed, generator speed, and generator power obtained at multiple predetermined time points through a multi-scale neighborhood feature extraction module to obtain multi-scale wind speed feature vectors, multi-scale impeller speed feature vectors, multi-scale generator speed feature vectors, and multi-scale power feature vectors, and calculates the corresponding responsiveness estimates to obtain a first cointegration transfer matrix, a second cointegration transfer matrix, and a third cointegration transfer matrix. These are then arranged into a three-dimensional input tensor and corrected by each feature matrix along the channel dimension through a convolutional neural network to obtain a corrected classification feature map. A classification result indicating whether the status of the wind turbine is normal can then be obtained through a classifier. In this way, the correlation characteristics between external environmental factors and internal factors of offshore wind turbines are mined to improve the accuracy and reliability of detection and judgment.

[0088] As described above, the wind turbine condition monitoring system 100 based on cointegration analysis according to an embodiment of the present application can be implemented in various terminal devices, such as a server with wind turbine condition monitoring based on cointegration analysis. In one example, the wind turbine condition monitoring system 100 based on cointegration analysis according to an embodiment of the present application can be integrated into the terminal device as a software module and / or hardware module. For example, the wind turbine condition monitoring system 100 based on cointegration analysis can be a software module in the operating system of the terminal device, or can be an application developed for the terminal device; of course, the wind turbine condition monitoring system 100 based on cointegration analysis can also be one of the many hardware modules of the terminal device.

[0089] Alternatively, in another example, the wind turbine condition monitoring system 100 based on cointegration analysis and the terminal device may also be separate devices, and the wind turbine condition monitoring system 100 based on cointegration analysis may be connected to the terminal device via a wired and / or wireless network and transmit interactive information in accordance with an agreed data format.

[0090] Exemplary Method

[0091] According to another aspect of the present application, a method for monitoring the status of a wind turbine generator based on cointegration analysis is also provided. Figure 7 As shown, the wind turbine condition monitoring method based on cointegration analysis according to the embodiment of the present application includes the following steps: S110, obtaining wind speed, impeller speed value, generator speed and generator power at multiple predetermined time points; S120, arranging the wind speed, impeller speed value, generator speed and generator power at the multiple predetermined time points into a wind speed input vector, an impeller speed input vector, a generator speed input vector and a power input vector respectively; S130, inputting the wind speed input vector, the impeller speed input vector, the generator speed input vector and the power input vector into a multi-scale neighborhood feature extraction module respectively to obtain a multi-scale wind speed feature vector, a multi-scale impeller speed feature vector, a multi-scale generator speed feature vector and a multi-scale power feature vector; S140, calculating the multi-scale impeller speed feature vector, the multi-scale generator speed feature vector and the multi-scale power feature vector respectively. The responsiveness of the multi-scale power eigenvector relative to the multi-scale wind speed eigenvector is estimated to obtain a first cointegration transfer matrix, a second cointegration transfer matrix and a third cointegration transfer matrix; S150, the first cointegration transfer matrix, the second cointegration transfer matrix and the third cointegration transfer matrix are arranged into a three-dimensional input tensor and then passed through adjacent layers using a convolutional neural network with mutually transposed convolution kernels to obtain a classification feature map; S160, based on a reference eigenvector, each eigenmatrix along the channel dimension of the classification feature map is corrected to obtain a corrected classification feature map, wherein the reference eigenvector is a channel eigenvector obtained by calculating the global mean of each eigenmatrix along the channel dimension of the classification feature map; and, S170, the corrected classification feature map is passed through a classifier to obtain a classification result, and the classification result is used to indicate whether the status of the wind turbine is normal.

[0092] In summary, the wind turbine state monitoring method based on cointegration analysis according to the embodiments of the present application is illustrated, which processes the wind speed, impeller speed value, generator speed and generator power obtained at multiple predetermined time points through a multi-scale neighborhood feature extraction module to obtain a multi-scale wind speed feature vector, a multi-scale impeller speed feature vector, a multi-scale generator speed feature vector and a multi-scale power feature vector, and calculates the corresponding responsiveness estimates to obtain a first cointegration transfer matrix, a second cointegration transfer matrix and a third cointegration transfer matrix, then arranges them into a three-dimensional input tensor, corrects each feature matrix along the channel dimension after the convolutional neural network to obtain a corrected classification feature map, and then uses a classifier to obtain a classification result representing whether the state of the wind turbine is normal. In this way, the correlation characteristics of external environmental factors and internal factors of offshore wind turbines are mined, and the detection accuracy and reliability are improved.

Claims

1. A wind turbine condition monitoring system based on cointegration analysis, characterized in that: include: A data acquisition and monitoring control module is used to obtain wind speed, impeller speed, generator speed and generator power at multiple predetermined time points; a vectorization module, configured to arrange the wind speed, impeller speed value, generator speed, and generator power at the plurality of predetermined time points into a wind speed input vector, an impeller speed input vector, a generator speed input vector, and a power input vector, respectively; a multi-scale time series feature extraction module, configured to input the wind speed input vector, the impeller speed input vector, the generator speed input vector, and the power input vector into a multi-scale neighborhood feature extraction module, respectively, perform one-dimensional convolution encoding on the wind speed input vector, the impeller speed input vector, the generator speed input vector, and the power input vector using a convolution layer with one-dimensional convolution kernels of different scales of the multi-scale neighborhood feature extraction module, and then cascade the obtained feature vectors corresponding to the two one-dimensional convolution kernels of different scales to obtain a multi-scale wind speed feature vector, a multi-scale impeller speed feature vector, a multi-scale generator speed feature vector, and a multi-scale power feature vector; a cointegration analysis module, configured to respectively calculate responsiveness estimates of the multi-scale impeller speed eigenvector, the multi-scale generator speed eigenvector, and the multi-scale power eigenvector relative to the multi-scale wind speed eigenvector to obtain a first cointegration transfer matrix, a second cointegration transfer matrix, and a third cointegration transfer matrix; a correlation feature extraction module, configured to arrange the first cointegration transfer matrix, the second cointegration transfer matrix, and the third cointegration transfer matrix into a three-dimensional input tensor and then pass the result through a convolutional neural network using mutually transposed convolution kernels in adjacent layers to obtain a classification feature map; a correction module, configured to correct each feature matrix of the classification feature map along the channel dimension based on a reference feature vector to obtain a corrected classification feature map, wherein the reference feature vector is a channel feature vector obtained by calculating the global mean of each feature matrix of the classification feature map along the channel dimension; as well as The state monitoring result generating module is used to pass the corrected classification feature map through a classifier to obtain a classification result, and the classification result is used to indicate whether the state of the wind turbine is normal.

2. The wind turbine condition monitoring system based on cointegration analysis according to claim 1 is characterized in that: The multi-scale temporal feature extraction module includes: a first convolutional encoding unit, configured to input the wind speed input vector, the impeller speed input vector, the generator speed input vector, and the power input vector, respectively, into a first convolutional layer of the multi-scale neighborhood feature extraction module to obtain a first neighborhood-scale wind speed-related feature vector, a first neighborhood-scale impeller speed-related feature vector, a first neighborhood-scale generator speed-related feature vector, and a first neighborhood-scale power-related feature vector, wherein the first convolutional layer has a first one-dimensional convolution kernel of a first length; a second convolutional encoding unit, configured to input the wind speed input vector, the impeller speed input vector, the generator speed input vector, and the power input vector, respectively, into a second convolutional layer of the multi-scale neighborhood feature extraction module to obtain a second neighborhood-scale wind speed-related feature vector, a second neighborhood-scale impeller speed-related feature vector, a second neighborhood-scale generator speed-related feature vector, and a second neighborhood-scale power-related feature vector, wherein the second convolutional layer has a second one-dimensional convolution kernel of a second length, and the first length is different from the second length; and a cascading unit, configured to concatenate the first neighborhood-scale wind speed association eigenvector and the second neighborhood-scale wind speed association eigenvector to obtain the multi-scale wind speed eigenvector, concatenate the first neighborhood-scale impeller speed association eigenvector and the second neighborhood-scale impeller speed association eigenvector to obtain the multi-scale impeller speed eigenvector, concatenate the first neighborhood-scale generator speed association eigenvector and the second neighborhood-scale generator speed association eigenvector to obtain the multi-scale generator speed eigenvector, and concatenate the first neighborhood-scale power association eigenvector and the second neighborhood-scale power association eigenvector to obtain the multi-scale power eigenvector.

3. The wind turbine condition monitoring system based on cointegration analysis according to claim 2 is characterized in that: The first convolutional encoding unit is further configured to: use the first convolutional layer of the multi-scale neighborhood feature extraction module to perform one-dimensional convolutional encoding on the wind speed input vector, the impeller speed input vector, the generator speed input vector, and the power input vector respectively using the following formula to obtain the first neighborhood-scale wind speed associated feature vector, the first neighborhood-scale impeller speed associated feature vector, the first neighborhood-scale generator speed associated feature vector, and the first neighborhood-scale power associated feature vector; Wherein, the formula is: Wherein, a is the width of the first convolution kernel in the x direction, E(a) is the first convolution kernel parameter vector, G(xa) is the local vector matrix operated with the convolution kernel function, w is the size of the first convolution kernel, and X represents the wind speed input vector, the impeller speed input vector, the generator speed input vector, and the power input vector.

4. The wind turbine condition monitoring system based on cointegration analysis according to claim 3 is characterized in that: The second convolutional encoding unit is further configured to: use the second convolutional layer of the multi-scale neighborhood feature extraction module to perform one-dimensional convolutional encoding on the wind speed input vector, the impeller speed input vector, the generator speed input vector, and the power input vector respectively using the following formula to obtain the second neighborhood-scale wind speed associated feature vector, the second neighborhood-scale impeller speed associated feature vector, the second neighborhood-scale generator speed associated feature vector, and the second neighborhood-scale power associated feature vector; Wherein, the formula is: Wherein, b is the width of the second convolution kernel in the x direction, F(b) is the second convolution kernel parameter vector, G(xb) is the local vector matrix operated with the convolution kernel function, m is the size of the second convolution kernel, and X represents the wind speed input vector, the impeller speed input vector, the generator speed input vector, and the power input vector.

5. The wind turbine condition monitoring system based on cointegration analysis according to claim 4 is characterized in that: The cointegration analysis module is further configured to calculate the responsiveness estimates of the multi-scale impeller speed eigenvector, the multi-scale generator speed eigenvector, and the multi-scale power eigenvector relative to the multi-scale wind speed eigenvector using the following formulas to obtain the first cointegration transfer matrix, the second cointegration transfer matrix, and the third cointegration transfer matrix; Wherein, the formula is: M i =M j *M1 Among them, M i They represent the multi-scale impeller speed eigenvector, the multi-scale generator speed eigenvector and the multi-scale power eigenvector respectively, M1 represents the multi-scale wind speed eigenvector, M j represent the first cointegration transfer matrix, the second cointegration transfer matrix and the third cointegration transfer matrix respectively.

6. The wind turbine condition monitoring system based on cointegration analysis according to claim 5 is characterized in that: The associated feature extraction module includes: A shallow feature map extraction unit, configured to extract a shallow feature map from the Mth layer of the convolutional neural network using mutually transposed convolution kernels of the adjacent layers, where M is an even number; a deep feature map extraction unit, configured to extract a deep feature map from an Nth layer of a convolutional neural network using mutually transposed convolution kernels in adjacent layers, wherein N is an even number and is greater than 2 times M; and A feature map fusion unit is used to fuse the shallow feature map and the deep feature map to generate the classification feature map.

7. The wind turbine condition monitoring system based on cointegration analysis according to claim 6, characterized in that: The correction module includes: A reference feature vector generating unit, configured to calculate a global mean of each feature matrix along a channel dimension of the classification feature map to obtain the reference feature vector; an optimization factor calculation unit, configured to calculate an optimization factor corresponding to an eigenvalue at each position in the reference eigenvector based on a mean and a variance of a set of eigenvalues ​​at all positions in the reference eigenvector; The correction unit is used to weight the corresponding feature matrix in the classification feature map using the optimization factor corresponding to the eigenvalue at each position in the reference eigenvector as a weight to obtain the corrected classification feature map.

8. The wind turbine condition monitoring system based on cointegration analysis according to claim 7 is characterized in that: The optimization factor calculation unit is further configured to calculate the optimization factor corresponding to the eigenvalue at each position in the reference eigenvector using the following formula based on the mean and variance of the eigenvalue set at all positions in the reference eigenvector; Wherein, the formula is: where v i represents the eigenvalue of the i-th position of the reference eigenvector, and μ and σ respectively represent the mean and variance of the eigenvalue set of all positions in the reference eigenvector, and exp(-σ) represents the calculation of the natural exponential function value raised to the negative value of the variance.

9. The wind turbine condition monitoring system based on cointegration analysis according to claim 8, characterized in that: The state monitoring result generation module is further configured to: the classifier processes the corrected classification feature map using the following formula to generate a classification result, wherein the formula is: softmax{(W n ,B n ):…:(W1,B1)|Project(F)}, where Project(F) represents projecting the corrected classification feature map into a vector, W1 to W n is the weight matrix of each fully connected layer, B1 to B n Represents the bias matrix of each fully connected layer.

10. A method for monitoring the condition of a wind turbine generator based on cointegration analysis, using the wind turbine generator condition monitoring system based on cointegration analysis according to any one of claims 1 to 9, characterized in that: include: Obtain wind speed, impeller speed, generator speed and generator power at multiple predetermined time points; Arranging the wind speed, impeller speed value, generator speed and generator power at the plurality of predetermined time points into a wind speed input vector, an impeller speed input vector, a generator speed input vector and a power input vector, respectively; Inputting the wind speed input vector, the impeller speed input vector, the generator speed input vector and the power input vector into a multi-scale neighborhood feature extraction module respectively to obtain a multi-scale wind speed feature vector, a multi-scale impeller speed feature vector, a multi-scale generator speed feature vector and a multi-scale power feature vector; respectively calculating responsiveness estimates of the multi-scale impeller speed eigenvector, the multi-scale generator speed eigenvector, and the multi-scale power eigenvector relative to the multi-scale wind speed eigenvector to obtain a first cointegration transfer matrix, a second cointegration transfer matrix, and a third cointegration transfer matrix; Arranging the first cointegration transfer matrix, the second cointegration transfer matrix, and the third cointegration transfer matrix into a three-dimensional input tensor and then passing it through a convolutional neural network with mutually transposed convolution kernels in adjacent layers to obtain a classification feature map; Based on a reference eigenvector, correcting each feature matrix along the channel dimension of the classification feature map to obtain a corrected classification feature map, wherein the reference eigenvector is a channel eigenvector obtained by calculating the global mean of each feature matrix along the channel dimension of the classification feature map; as well as The corrected classification feature map is passed through a classifier to obtain a classification result, and the classification result is used to indicate whether the state of the wind turbine is normal.

Images