Cluster balancing representation and spatiotemporal coordinated hierarchical monitoring model construction method for wind turbine

Abstract

The application discloses a kind of cluster equalization characterization and space-time coordination hierarchical monitoring model construction methods for wind turbine unit.The method constructs hierarchical monitoring model from the angle of space-time coordination.First, a single-layer monitoring model is established, and the stationary features and non-stationary features of the wind turbine time series signal are extracted respectively. The variance fluctuation of the stationary features and the dynamic change of the non-stationary features are integrated to design monitoring statistics to monitor the operating state of the single-layer. Then, a cluster layer monitoring model is established by fusing different single-body spaces. The inherent equalization features and incentive equalization features are obtained by cluster feature engineering. The static deviation and fluctuation variation of the features are used to design monitoring statistics to monitor the operating state of the cluster layer. The application can mine the characteristics of the wind turbine cluster, improve the sensitivity of the wind turbine monitoring object, and provide practical help for the health monitoring and equipment operation and maintenance of the wind turbine unit.

Description

Technical Field

[0001] This invention belongs to the field of health monitoring and fault early warning of wind turbine cluster equipment. In particular, it proposes a method for constructing a cluster equilibrium representation and spatiotemporal collaborative hierarchical monitoring model for wind turbine units. A hierarchical monitoring model is constructed to monitor the operating status of wind turbine equipment from both individual and cluster perspectives. The inherent equilibrium representation and excitation equilibrium representation of the cluster are obtained from the individual units by utilizing cluster spatiotemporal state feature engineering. An online monitoring strategy is designed to provide real-time early warning of abnormal states of wind turbine equipment. Background Technology

[0002] With the continuous growth of global population and economy, the world's energy demand is gradually increasing. Traditional fossil fuels not only cannot meet the needs of sustainable development but also harm the environment and climate. Therefore, the transformation of energy production and consumption patterns and energy structure has become a key task for the next two to three decades, and the development of renewable energy sources such as wind and solar energy has received widespread attention. Wind energy, as an energy generated by air movement, has broad development prospects. As the basic energy conversion equipment in the wind energy industry, wind turbine generators rely on wind blowing the blades to rotate, which in turn drives the motor coils to cut the magnetic field and generate electricity. This process involves multiple sub-equipment, and the frequently changing operating conditions and strongly coupled sensing points make it difficult to detect unit faults in a timely manner, bringing great difficulties to equipment operation and maintenance. Due to the power generation principle and structural design of wind turbines, the internal working environment of the unit is usually in a changing environment, making it difficult to accurately detect some subtle faults. Existing monitoring methods for wind turbine equipment only assess the health status of the individual wind turbine and do not explore the non-stationary characteristics during wind turbine operation in depth. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention designs a hierarchical monitoring model that analyzes sample characteristics from both individual and cluster perspectives. Through cluster spatiotemporal state feature engineering, it extracts a comprehensive and balanced cluster representation from individual sample characteristics, enabling fine-grained assessment of abnormal states. Simultaneously, considering the non-stationarity of wind turbine operating states, it finely distinguishes between stationary and non-stationary states from a time dimension, completing a multi-state health perception process.

[0004] The purpose of this invention is to address the shortcomings of existing technologies by providing a method for cluster equilibrium characterization and spatiotemporal collaborative hierarchical monitoring of wind turbine units.

[0005] The objective of this invention is achieved through the following technical solution:

[0006] A method for constructing a cluster equilibrium representation and spatiotemporal collaborative hierarchical monitoring model for wind turbine generators includes a single-unit layer monitoring model and a cluster layer monitoring model. The specific construction method is as follows:

[0007] (1) Obtain the three-dimensional data X = [x1,...,x2] of all wind turbines deployed in a wind turbine cluster during normal operation. i ,…,x I ], where x i This represents the two-dimensional data of the i-th wind turbine during normal operation. The two-dimensional data consists of J variable data collected at K time points during the normal operation of the corresponding wind turbine, and I represents the number of wind turbines.

[0008] (2) Based on the normal operation data of the corresponding wind turbine equipment, I individual-level monitoring models are constructed and trained using stationary subspace analysis for each of the I individual wind turbine equipment; the individual-level monitoring models will use the two-dimensional data x of the corresponding wind turbine equipment. i Transform into latent variable s i =[s st,i ,s nst,i ] T , where s i Let s represent the latent variable of the i-th wind turbine unit. st,i Let s represent the stationary latent characteristic variable of the i-th wind turbine unit. nst,i Let i represent the non-stationary latent variable of the i-th individual; then, design two individual monitoring statistics based on the I-individual layer monitoring model. and Calculate the corresponding control limits, where It is a single-unit monitoring statistic built based on the monitoring model of each single-unit layer. The mean value is used to calculate the stationary characteristic state of the individual wind turbine operating data. It is a single-unit monitoring statistic built based on the monitoring model of each single-unit layer. The mean value is used to calculate the non-stationary characteristic state of individual wind turbine operating data;

[0009] (3) Compare the latent variables s containing the self-stationary characteristics of the corresponding wind turbine equipment. st,i Non-stationary latent variable s nst,i The difference before and after is extracted as the inherent equilibrium characterization (CEF) of the corresponding wind turbine unit. st,i and incentive equilibrium characterization CEF nst,i Constructing a cluster equilibrium representation (CEF) for the corresponding individual wind turbine equipment i =[cef st,i ,cef nst,i The cluster-level monitoring model is constructed by taking the average of the cluster equilibrium representations of all individual wind turbine units and constructing a comprehensive cluster equilibrium representation. Based on this model, two cluster monitoring statistics are then designed. and Calculate the corresponding control limits, where Used to calculate the operating status of the inherent balance attribute performance of wind turbine equipment cluster data. Used to calculate the operational status of the data incentive balance attribute performance of wind turbine equipment clusters.

[0010] Furthermore, the single-layer monitoring model will correspond to the two-dimensional data x of the wind turbine equipment. i Transform into latent variable s i Specifically:

[0011] s i =W i x i =[W s i x i W n i x i ] T

[0012] Among them, W i W is the invertible linear transformation matrix of the i-th wind turbine unit. s i For stationary characteristic latent variables s st,i The linear transformation matrix, W n i For non-stationary latent variable s nst,i The linear transformation matrix;

[0013] W i It is obtained by solving based on the normal operating data of the corresponding wind turbine equipment, specifically...

[0014] The two-dimensional data x of the i-th wind turbine unit i Divide the time into non-overlapping time slices and calculate the mean of each time slice j. Covariance

[0015] Taking advantage of the property that the mean and covariance of stationary data remain unchanged, a linear transformation process is used to obtain the mean of the stationary latent variable. Covariance

[0016]

[0017] According to the definition of stationarity, the sum of the differences in mean and covariance at each time slice should be minimized. To ensure that different stationary features carry different information, constraints are imposed. Where I is the identity matrix, W i The problem is transformed into a minimization problem.

[0018]

[0019] Where n is the total number of time slices, D KLThis represents solving for the KL divergence of two distributions. The mean is covariance is The Gaussian distribution is N(0,I), which represents a Gaussian distribution with mean 0 and covariance I.

[0020] Furthermore, the design includes two individual monitoring statistics. and The specific calculation of the corresponding control limits is as follows:

[0021]

[0022]

[0023]

[0024]

[0025] Where Δs nst,i s nst,i The first-order difference in the time dimension, Λ represents s. nst,i The eigenvalues ​​of the covariance matrix are the diagonal matrix. and These are the stationary feature monitoring statistics and the non-stationary feature monitoring statistics constructed based on the i-th single-layer monitoring model, respectively.

[0026] The control limits corresponding to the statistics of the i-th wind turbine equipment are obtained by using kernel density estimation (KDE). and The control limit for individual monitoring statistics is the average of the control limits for each wind turbine device.

[0027]

[0028] Furthermore, the inherent equilibrium characterization of the corresponding wind turbine unit is cef. st,i and incentive equilibrium characterization CEF nst,i It is expressed as follows:

[0029] CEF st,i =[Stat 1,st,i ,Stat 2,st,i ]

[0030] CEF nst,i =[Stat 1,nst,i ,Stat 2,nst,i ]

[0031] Among them, Stat 1,st,i and Stat 2,st,i These represent the latent variables s that include and do not include the stationary characteristic of the i-th wind turbine unit. st,iThe difference in mean and variance before and after, Stat 1,nst,i and Stat 2,nst,i These represent the latent variables s that include and do not include the non-stationary characteristics of the i-th wind turbine unit itself. nst,i The difference in mean and variance before and after; specifically,

[0032] Stat 1,st,i =mean st -mean st,i

[0033] Stat 2,st,i =var st -var st,i

[0034] Stat 1,nst,i =mean nst -mean nst,i

[0035] Stat 2,nst,i =var nst -var nst,i

[0036] Where ⊙ represents the Hamiltonian product, mean st mean st,i mean nst mean nst,i var st var st,i var nst var nst,i The expressions are respectively

[0037]

[0038] Furthermore, the cluster layer monitoring model is specifically as follows:

[0039] The cluster-level monitoring model is constructed from the cluster comprehensive equilibrium representation (CEF), which is obtained by averaging the equilibrium representations of each individual unit.

[0040]

[0041] Among them, cef st This represents the inherent equilibrium characteristic of a wind turbine cluster, cef nst This represents the excitation equilibrium characteristic of a wind turbine cluster.

[0042] Furthermore, two cluster monitoring statistics are designed based on the cluster-layer monitoring model. and The specific calculations for the corresponding control limits are as follows:

[0043] After normalizing the cluster equilibrium representation according to the time dimension, the statistic is calculated as follows:

[0044]

[0045]

[0046] Among them Λ st and Λ nst They represent cef respectively st and CEF nst The covariance matrix is ​​a diagonal matrix of eigenvalues.

[0047] The control limits corresponding to the statistic are obtained by estimating the KDE using kernel density. and

[0048] A method for cluster equilibrium representation and spatiotemporal coordinated hierarchical monitoring of wind turbines is implemented based on the cluster equilibrium representation and spatiotemporal coordinated hierarchical monitoring model construction method for wind turbines, specifically as follows:

[0049] The system acquires real-time operating data of all wind turbines in a wind turbine cluster. Based on the aforementioned method for constructing a cluster equilibrium representation and spatiotemporal collaborative hierarchical monitoring model for wind turbines, it constructs a cluster equilibrium representation and spatiotemporal collaborative hierarchical monitoring model for wind turbines. The system calculates individual monitoring statistics and cluster monitoring statistics for online samples and determines whether the sample was collected during a period of equipment failure based on the corresponding control limits, thereby monitoring the operating status of the wind turbine cluster.

[0050] Furthermore, based on the corresponding control limits, it is determined whether the sample was collected during the period of equipment failure, specifically as follows:

[0051] If any of the calculated monitoring statistics exceeds the corresponding control limit, it indicates that the wind turbine equipment has malfunctioned at that moment; if any of the calculated individual monitoring statistics exceeds its control limit, it indicates that the individual wind turbine equipment is in an abnormal operating state; if any of the calculated cluster monitoring statistics exceeds its control limit, it indicates that the wind turbine cluster is in an abnormal operating state.

[0052] Compared with existing technologies, the beneficial effects of this invention are as follows: A hierarchical monitoring model is designed, analyzing sample characteristics from both individual and cluster perspectives. Through cluster spatiotemporal state feature engineering, a comprehensive and balanced cluster representation is extracted from individual sample characteristics, achieving fine-grained anomaly assessment. Simultaneously, addressing the non-stationarity of wind turbine operating states, a refined distinction is made between stationary and non-stationary states from a time dimension, completing a multi-state health perception process. Compared to existing wind turbine anomaly monitoring methods, this invention establishes a hierarchical state analysis perspective for the first time, effectively improving the fault detection rate in monitoring the operating status of multiple wind turbine equipment groups operating in parallel within the same area. This provides practical assistance for the refined health management of large-scale wind turbine equipment and the operation and maintenance of wind turbine cluster equipment groups. Attached Figure Description

[0053] Figure 1 A flowchart of a method for constructing a cluster equilibrium representation and spatiotemporal collaborative hierarchical monitoring model for wind turbines (left) and a method for cluster equilibrium representation and spatiotemporal collaborative hierarchical monitoring of wind turbines (right);

[0054] Figure 2 The graph shows the test results of fault detection on online operating data for the method proposed in this invention. The dashed line represents the natural logarithm of the control limit, and the solid line represents the natural logarithm of the statistical calculation results of the test set samples.

[0055] Figure 3 The graph shows the test results of fault detection using the stationary subspace analysis method on online running data. The dashed line represents the natural logarithm of the control limits, and the solid line represents the natural logarithm of the statistical calculation results of the test set samples.

[0056] Figure 4 The graph shows the test results of the slow feature analysis method for fault detection on online running data. The dashed line represents the natural logarithm of the control limits, and the solid line represents the natural logarithm of the statistical calculation results of the test set samples. Detailed Implementation

[0057] This invention provides a method for constructing a cluster equilibrium representation and spatiotemporal coordinated hierarchical monitoring model for wind turbines. The model includes a single-unit layer monitoring model and a cluster layer monitoring model. The construction method includes:

[0058] Obtain the three-dimensional data X = [x1, ..., x2] of all wind turbines deployed in a wind turbine cluster during normal operation. i ,…,x I ], where x i Let I represent the two-dimensional data of the i-th wind turbine during normal operation, and let I represent the number of wind turbines. Let the data matrix of the i-th wind turbine (containing J variables) at K moments during normal operation be denoted as...

[0059] A single-unit monitoring model is constructed, with I sub-models established for each I wind turbine unit. Each sub-model is trained using the normal operating data of its corresponding wind turbine unit. Based on the non-stationarity of the wind turbine unit's operating data, Stationary Subspace Analysis (SSA) is used to transform the original unit features x... i Transform into latent variable s i =[s st,i ,s nst,i ], where s i Let s represent the latent variable of the i-th wind turbine unit. st,i Let s represent the stationary latent characteristic variable of the i-th wind turbine unit. nst,i Let the non-stationary latent characteristic variable of the i-th wind turbine unit be represented. Design two unit monitoring statistics. and in Used to calculate the stationary characteristic state of individual wind turbine equipment operating data. Used to calculate the non-stationary characteristic state of individual wind turbine equipment operating data. Calculates control limits for monitoring statistics of two individual units. and in Indicates the individual statistics of wind turbine equipment Corresponding control limits, Indicates the individual statistics of wind turbine equipment Corresponding control limits.

[0060] Constructing the spatiotemporal characteristics of the wind turbine cluster, and using the transformation in the individual unit monitoring model to obtain the latent variables s of the individual wind turbine equipment. i Extract the equilibrium representation of the wind turbine cluster, where i represents the sequence number of the individual wind turbine unit. Compare the latent variables s containing the stationary characteristics of the corresponding individual wind turbine unit over time. st,i Non-stationary latent variable s nst,i The differences before and after are used to extract the inherent equilibrium attribute and the incentivized equilibrium attribute of the individual unit. In the spatial dimension, the statistical contribution feature Stat represents the impact of the individual unit on the cluster, and cluster equilibrium representations belonging to different units are obtained. For the i-th wind turbine unit, the extracted cluster equilibrium representation is cef. i =[cef st,i ,cef nst,i ], where cef st,i This represents the stationary latent variable s of the i-th wind turbine unit. st,i Extracted inherent equilibrium properties, CEF nst,i This represents the non-stationary latent variable s of the i-th wind turbine unit. nst,i Extracted incentive equilibrium attributes.

[0061] A cluster-level monitoring model is constructed by averaging the cluster equilibrium representations of all individual wind turbine units. This model is trained using data X from all wind turbine units, and the equilibrium representation of the wind turbine cluster is obtained based on the spatiotemporal characteristics of the cluster. Two cluster monitoring statistics are designed. and in Used to calculate the operating status of the inherent balance attribute performance of wind turbine equipment cluster data. This is used to calculate the operational status of the data incentive balancing attributes of the wind turbine equipment cluster. It also calculates the control limits of monitoring statistics for the two clusters. and in This represents the statistical value of the inherent equilibrium property of the wind turbine cluster. Corresponding control limits, This represents the statistics of the excitation equilibrium attribute of the wind turbine cluster. Corresponding control limits.

[0062] Based on the hierarchical monitoring model of individual unit layer to cluster layer, an online monitoring strategy is designed to acquire the operating data of the wind turbine group in real time, calculate the monitoring statistics of online samples, and determine whether the sample was collected during the equipment failure period according to the control limit. The operation status of the wind turbine equipment group can be monitored and early warning can be achieved.

[0063] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. First, historical normal operation data of the wind turbines is acquired and preprocessed using normalization. Then, a single-unit monitoring model is constructed, and the parameters of the single-unit monitoring model are obtained by training with historical data. Next, cluster spatiotemporal state feature engineering is used to extract the cluster comprehensive equilibrium representation from the single-unit sample features in the single-unit monitoring model, constructing a wind turbine cluster-level monitoring model, and the model parameters are obtained by training with the cluster comprehensive equilibrium representation. Finally, the single-unit-cluster hierarchical monitoring model is deployed in the online monitoring process to receive online operation data of the wind turbine units, and anomaly warning results are obtained through monitoring strategy judgment. This embodiment uses a cluster of 5 wind turbine units as an example to verify the effectiveness of this method, specifically including the following steps:

[0064] Step 1: Obtain data on the normal operation of all wind turbines in a wind turbine cluster, X = [x1,...,x...]. i ,...,x I ], where x i Let I represent the data of the i-th wind turbine during normal operation, and let I represent the number of wind turbines. The data matrix of the i-th wind turbine (containing J variables) at K moments during normal operation is denoted as... In this embodiment, the wind turbine cluster contains I = 5 wind turbines; each wind turbine contains J = 36 variables, including wind speed, elevation angle, etc.; the sampling period is 10s, and there are a total of K = 15000 samples.

[0065] Finally, the data from the normal operation of each wind turbine is normalized and used for feature extraction in the subsequent hierarchical monitoring model.

[0066] Step 2: Construct a single-unit layer monitoring model. For each individual wind turbine unit, establish one single-unit layer monitoring model. Each single-unit layer monitoring model is trained using the normal operating data of the corresponding wind turbine unit. The specific construction process is as follows:

[0067] Based on the non-stationarity of individual wind turbine operating data, stationary subspace analysis (SSA) is used to normalize the historical normal operating data x of the i-th wind turbine. i Transform into latent variable s i =[s st,i ,s nst,i ] T ,

[0068] s i =W i x i =[W s i x i W n i x i ] T

[0069] Among them, W i W is an invertible linear transformation matrix. s i For stationary features s st,i The linear transformation matrix, W n i Non-stationary feature s nst,i The linear transformation matrix.

[0070] Solve for W i The process specifically involves processing the operating data x of the i-th wind turbine unit. i Divide the data into non-overlapping segments based on time and define them as Γ1, Γ2, ..., Γ n n is the total number of time slices. The mean of the j-th time slice. Covariance It can be obtained using runtime data.

[0071]

[0072] Among them, T j For time slices Γj The number of samples included. x i (t) represents the t-th sample of the i-th wind turbine unit, which contains the variable data sampled at that time.

[0073] Taking advantage of the property that the mean and covariance of stationary data remain unchanged, a linear transformation process is used to obtain the mean of the stationary latent variable. Covariance

[0074]

[0075] According to the definition of stationarity, the sum of the differences in mean and covariance at each time slice should be minimized. To ensure that different stationary features carry different information, constraints are imposed. Where I is the identity matrix, W i The problem is transformed into a minimization problem.

[0076]

[0077] Where D KL This represents solving for the KL divergence of two distributions. The mean is covariance is The Gaussian distribution is N(0,I), which represents a Gaussian distribution with mean 0 and covariance I.

[0078] The calculation process for the individual layer monitoring statistics of the k-th sample in the historical normal data of the i-th wind turbine equipment is as follows:

[0079]

[0080]

[0081]

[0082]

[0083] Where Δs nst,i (k) represents s nst,i After the first-order difference in the time dimension, the k-th sample, Λ represents s. nst,i The eigenvalues ​​of the covariance matrix of (k) are used as the result after training the model. and These are the monitoring statistics of stationary characteristics and non-stationary characteristics of the individual layer for the i-th wind turbine equipment and the k-th sample, respectively.

[0084] Based on 15,000 statistics derived from the historical data of the i-th wind turbine unit, the corresponding control limits are estimated using kernel density estimation (KDE). and Set the control limit confidence level to 90%, and set the control limit for individual monitoring statistics to the average of the control limits for all devices.

[0085]

[0086]

[0087] in and These are the individual layer monitoring statistics. and The control limits are used as the result after training the model.

[0088] Step 3: Construct the spatiotemporal characteristics of the cluster, and utilize the latent variable s of the i-th wind turbine unit obtained from the transformation in the unit-level monitoring model. i Extracting the comprehensive equilibrium characterization of the wind turbine equipment cluster, specifically:

[0089] Selecting the statistical contribution feature Stat as the mean and variance, the inherent equilibrium representation of the contribution of the i-th wind turbine to the cluster is cef. st,i and incentive equilibrium characterization CEF nst,i The solution is as follows:

[0090] CEF st,i =[Stat 1,st,i ,Stat 2,st,i ]

[0091] CEF nst,i =[Stat 1,nst,i ,Stat 2,nst,i ]

[0092] Among them, Stat 1,st,i and Stat 2,st,i Let Stat represent the statistical contribution characteristics of the i-th wind turbine unit obtained from the mean and variance of the stationary latent variables, respectively. 1,nst,i and Stat 2,nst,i These represent the statistical contribution characteristics of the i-th wind turbine unit, obtained from the mean and variance of the non-stationary latent variables, respectively. Specifically,

[0093] Stat 1,st,i =mean st -mean st,i

[0094] Stat 2,st,i =var st -var st,i

[0095] Stat 1,nst,i =mean nst -mean nst,i

[0096] Stat 2,nst,i =var nst -var nst,i

[0097] Where ⊙ represents the Hamiltonian product, mean st mean st,i mean nst mean nst,i var st var st,i var nst var nst,i The expressions are respectively

[0098]

[0099]

[0100]

[0101]

[0102]

[0103]

[0104] The contribution of the i-th wind turbine to the equilibrium representation of the cluster can be represented by the inherent equilibrium representation and the incentivized equilibrium representation, i.e., cef i =[cef st,i ,cef nst,i The cluster's overall equilibrium representation (CEF) is obtained by averaging the contributions of each individual unit to the cluster's equilibrium representation.

[0105]

[0106] Among them, cef st This represents the inherent equilibrium characteristic of a wind turbine cluster, cef nst This represents the excitation equilibrium characteristic of a wind turbine cluster.

[0107] Step 4: Construct a cluster-level monitoring model, and obtain a comprehensive equilibrium representation calculation and training model for the wind turbine cluster based on the spatiotemporal state characteristics of the cluster. Specifically:

[0108] The cluster comprehensive equilibrium representation (CEF) is used as the constructed cluster-level monitoring model;

[0109] After normalizing the cluster comprehensive equilibrium representation according to the time dimension, the cluster-level monitoring statistic for the k-th sample is calculated as follows:

[0110]

[0111]

[0112] Among them Λ st and Λ nst They represent cef respectively st (k) and cef nst (k) The eigenvalues ​​of the covariance matrix are used as the result after training the model. and These are the cluster-level inherent equilibrium attribute statistics and the incentive equilibrium attribute statistics for the k-th sample of the wind turbine equipment, respectively.

[0113] The control limits corresponding to the statistic are obtained by estimating the KDE using kernel density. and The control limit confidence level was set to 90%, which was used as the result after training the model. This yielded the cluster equilibrium representation and spatiotemporal collaborative hierarchical monitoring model for wind turbines constructed in this invention.

[0114] Based on a hierarchical monitoring model of individual unit layer and cluster layer, a cluster equilibrium representation and spatiotemporal collaborative hierarchical monitoring method for wind turbine units is designed. Specifically, it involves: acquiring real-time operating data of the wind turbine group, calculating online samples to obtain monitoring statistics, and determining whether the sample was collected during a period of equipment failure based on control limits, thereby monitoring the operating status of the wind turbine group. Specifically:

[0115] Obtain online wind turbine operation data Where x new,i This represents the online operating data of the i-th wind turbine at the current moment. Based on the unit-cluster structure of the hierarchical monitoring model, the operating data is first normalized and then transformed into a latent variable s. new ,

[0116]

[0117]

[0118]

[0119] in Let m represent the stationary latent feature variables of the online data, and m represent the dimension of the stationary latent feature variables. For non-stationary latent variables of online data, and Let W represent the stationary and non-stationary latent characteristic variables of the online data of the i-th wind turbine unit, respectively. s i and W n i These are the linear transformation matrix parameters of the single-layer monitoring model after training.

[0120] Calculate the online data statistics for each of the I wind turbine units. The statistics for the i-th wind turbine unit in the single-unit monitoring model are calculated as follows:

[0121]

[0122]

[0123]

[0124]

[0125] in and These represent the stationary and non-stationary state statistics of the online data of the i-th wind turbine unit, respectively. and This represents the statistics calculated by the single-layer monitoring model based on online operational data. Λ represents the model parameters obtained during the training phase.

[0126] Then, cluster feature engineering is used to calculate the inherent equilibrium representation of the online data. Incentive Equilibrium Characterization and comprehensive equilibrium characterization of CEF new Calculate cluster-level monitoring statistics based on online data.

[0127]

[0128]

[0129] Among them Λ st and Λ nst These are the model parameters obtained during the training phase.

[0130] Finally, the control limits calculated during the training phase are... As control limits for the corresponding statistics, stratified comparisons are performed. If any monitoring statistic calculated based on the online operation data of the wind turbine unit exceeds the corresponding control limit, it indicates that the wind turbine equipment has malfunctioned at that moment; specifically, if any individual monitoring statistic exceeds its control limit, it indicates that the individual wind turbine is operating abnormally; if any cluster monitoring statistic exceeds its control limit, it indicates that the wind turbine cluster is operating abnormally.

[0131] The online operation test had a data sample size of 4000. The fifth wind turbine failed at time 2461. The monitoring results of the failure using the method proposed in this invention are as follows: Figure 2As shown in the figure. The dashed line in the figure represents the natural logarithm of the control limit, and the solid line represents the natural logarithm of the statistical calculation results of the test set sample. It can be found that when the wind turbine is abnormal, all the statistics of this method exceed the control limit after the fault occurs, and the abnormality is detected immediately after the fault occurs. In order to more clearly demonstrate the superiority of this invention in the anomaly detection task, stationary subspace analysis (SSA) (Blythe, DA, Von Bunau, P., Meinecke, FC, & Muller, KRF) and slow feature analysis (SFA) (C. Shang, B. Huang, F. Yang, et al. "Slow feature analysis for monitoring and diagnosis of control performance." J Process Control, vol. 39, pp. 21-34, 2016) are compared with the method of this invention. Neither the SSA nor SFA methods involve cluster-level analysis; therefore, only the test data from the fifth wind turbine was used for monitoring. The fault detection performance of the above methods on the test set is as follows: Figure 3 and Figure 4 As shown; each curve represents the result after taking the natural logarithm of the corresponding statistic and its control limits.

[0132] For a more intuitive comparison, Table 1 lists the false alarm rate (MAR) and false alarm rate (FAR) of different methods on the test set. The comparison shows that the method proposed in this invention significantly reduces both the MAR and FAR of different statistics compared to the comparative methods, demonstrating the superior performance of the method in wind turbine health monitoring.

[0133] Comparison of false alarm rate and false alarm rate of different methods on test data

[0134]

[0135] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for constructing a cluster equilibrium representation and spatiotemporal collaborative hierarchical monitoring model for wind turbine generators, characterized in that, The cluster equilibrium representation and spatiotemporal collaborative hierarchical monitoring model for wind turbines includes a single-unit layer monitoring model and a cluster layer monitoring model, and the construction method includes: Obtain the three-dimensional data X = [x1,...,x] of all wind turbines deployed in a wind turbine cluster during normal operation. i ,...,x I ], where x i This represents the two-dimensional data of the i-th wind turbine during normal operation. The two-dimensional data consists of J variable data collected at K time points during the normal operation of the corresponding wind turbine, and I represents the number of wind turbines. Based on the normal operation data of the corresponding wind turbine equipment, stationary subspace analysis is used to construct and train I individual-level monitoring models for each individual wind turbine equipment; the individual-level monitoring models will use the two-dimensional data x of the corresponding wind turbine equipment. i Transform into latent variable s i =[s st,i ,s nst,i ] T , where s i Let s represent the latent variable of the i-th wind turbine unit. st,i Let s represent the stationary latent characteristic variable of the i-th wind turbine unit. nst,i Let i represent the non-stationary latent variable of the i-th individual; then, design two individual monitoring statistics based on the I-individual layer monitoring model. and Calculate the corresponding control limits, where It is a single-unit monitoring statistic built based on the monitoring model of each single-unit layer. The mean value is used to calculate the stationary characteristic state of the individual wind turbine operating data. It is a single-unit monitoring statistic built based on the monitoring model of each single-unit layer. The mean value is used to calculate the non-stationary characteristic state of individual wind turbine operating data; Compare the latent variables s, which include the stationary characteristics of the corresponding individual wind turbine equipment. st,i Non-stationary latent variable s nst,i The difference before and after is extracted as the inherent equilibrium characterization (CEF) of the corresponding wind turbine unit. st,i and incentive equilibrium characterization CEF nst,i Constructing a cluster equilibrium representation (CEF) for the corresponding individual wind turbine equipment i =[cef st,i ,cef nst,i The cluster-level monitoring model is constructed by taking the average of the cluster equilibrium representations of all individual wind turbine units and constructing a comprehensive cluster equilibrium representation. Based on this model, two cluster monitoring statistics are then designed. and Calculate the corresponding control limits, where Used to calculate the operating status of the inherent balance attribute performance of wind turbine equipment cluster data. Used to calculate the operational status of the data incentive balance attribute performance of wind turbine equipment clusters.

2. The method according to claim 1, characterized in that, The single-layer monitoring model corresponds to the two-dimensional data x of the wind turbine equipment. i Transform into latent variable s i Specifically: s i =W i x i =[W s i x i ,W n i x i ] T Among them, W i W is the invertible linear transformation matrix of the i-th wind turbine unit. s i For stationary characteristic latent variables s st,i The linear transformation matrix, W n i For non-stationary latent variable s nst,i The linear transformation matrix; W i It is obtained by solving based on the normal operating data of the corresponding wind turbine equipment, specifically... The two-dimensional data x of the i-th wind turbine unit i Divide the time into non-overlapping time slices and calculate the mean of each time slice j. Covariance Taking advantage of the property that the mean and covariance of stationary data remain unchanged, a linear transformation process is used to obtain the mean of the stationary latent variable. Covariance According to the definition of stationarity, the sum of the differences in mean and covariance at each time slice should be minimized. To ensure that different stationary features carry different information, constraints are imposed. Where I is the identity matrix, W i The problem is transformed into a minimization problem. Where n is the total number of time slices, D KL This represents solving for the KL divergence of two distributions. The mean is covariance is The Gaussian distribution is N(0,I), which represents a Gaussian distribution with mean 0 and covariance I.

3. The method according to claim 1, characterized in that, The design includes two individual monitoring statistics. and The specific calculation of the corresponding control limits is as follows: Where Δs nst,i s nst,i The first-order difference in the time dimension, Λ represents s. nst,i The eigenvalues ​​of the covariance matrix are the diagonal matrix. and These are the stationary feature monitoring statistics and the non-stationary feature monitoring statistics constructed based on the i-th single-layer monitoring model, respectively. The control limits corresponding to the statistics of the i-th wind turbine equipment are obtained by using kernel density estimation (KDE). and The control limit for individual monitoring statistics is the average of the control limits for each wind turbine device.

4. The method according to claim 1, characterized in that, The inherent equilibrium characterization of the corresponding wind turbine unit (cef) st,i and incentive equilibrium characterization CEF nst,i It is expressed as follows: boss st,i =[State 1,st,i ,State 2,st,i ] boss nst,i =[State 1,nst,i ,State 2,nst,i ] Among them, Stat 1,st,i and Stat 2,st,i These represent the latent variables s that include and do not include the stationary characteristic of the i-th wind turbine unit. st,i The difference in mean and variance before and after, Stat 1,nst,i and Stat 2,nst,i These represent the latent variables s that include and do not include the non-stationary characteristics of the i-th wind turbine unit itself. nst,i The difference in mean and variance before and after; specifically, Stat 1,st,i =mean st -mean st,i State 2,st,i =where st -was st,i Stat 1,nst,i =mean nst -mean nst,i State 2,nst,i =where nst -was nst,i Where ⊙ represents the Hamiltonian product, mean st mean st,i mean nst mean nst,i var st var st,i var nst var nst,i The expressions are respectively 5. The method according to claim 1, characterized in that, The cluster layer monitoring model is specifically as follows: The cluster-level monitoring model is constructed from the cluster comprehensive equilibrium representation (CEF), which is obtained by averaging the equilibrium representations of each individual unit. Among them, cef st This represents the inherent equilibrium characteristic of a wind turbine cluster, cef nst This represents the excitation equilibrium characteristic of a wind turbine cluster.

6. The method according to claim 1, characterized in that, Two cluster monitoring statistics were designed based on the cluster layer monitoring model. and The specific calculations for the corresponding control limits are as follows: After normalizing the cluster equilibrium representation according to the time dimension, the statistic is calculated as follows: Among them Λ st and Λ nst They represent cef respectively st and CEF nst The covariance matrix is ​​a diagonal matrix of eigenvalues. The control limits corresponding to the statistic are obtained by estimating the KDE using kernel density. and 7. A method for cluster equilibrium characterization and spatiotemporal collaborative hierarchical monitoring of wind turbine generators, characterized in that, The implementation of the cluster equilibrium representation and spatiotemporal collaborative hierarchical monitoring model for wind turbines, based on the construction method of the cluster equilibrium representation and spatiotemporal collaborative hierarchical monitoring model for wind turbines according to any one of claims 1-6, is as follows: The system acquires real-time operating data of all wind turbines in a wind turbine cluster. Based on the cluster equilibrium representation and spatiotemporal collaborative hierarchical monitoring model for wind turbines constructed according to any one of claims 1-6, the system calculates the individual monitoring statistics and cluster monitoring statistics of online samples. It also determines whether the sample was collected during a period of equipment failure based on the corresponding control limit, thereby monitoring the operating status of the wind turbine cluster.

8. The method according to claim 7, characterized in that, Determine whether the sample was collected during the period of equipment failure based on the corresponding control limits, specifically: If any of the calculated monitoring statistics exceeds the corresponding control limit, it indicates that the wind turbine equipment has malfunctioned at that moment; if any of the calculated individual monitoring statistics exceeds its control limit, it indicates that the individual wind turbine equipment is in an abnormal operating state; if any of the calculated cluster monitoring statistics exceeds its control limit, it indicates that the wind turbine cluster is in an abnormal operating state.

Images