A wind turbine operating state monitoring method based on multi-source heterogeneous data

Abstract

The application discloses a wind turbine operation state monitoring method based on multi-source heterogeneous data, comprising: collecting SCADA multi-source heterogeneous data, and eliminating abnormal samples and standardizing; based on space rank, converting the collected non-normal data into approximate multivariate normal distribution; constructing an elastic net penalty likelihood function, and solving the mean vector estimation representing the SCADA sparse fault signal; based on the mean vector estimation, calculating the basic monitoring statistics and robust optimization, obtaining the robust monitoring statistics and upper control limit; calculating the real-time robust monitoring statistics, if exceeding the upper control limit, issuing a fault alarm; after triggering the fault alarm, according to the optimal penalty coefficient, backstepping the mean vector estimation, identifying and outputting the fault component information. The application solves the problems of complex SCADA data distribution, variable grouping correlation interference, sparse fault difficulty in identification, insufficient controlled samples and high multi-unit monitoring cost in the prior art.

Description

Technical Field

[0001] This invention relates to the field of wind turbine operation and maintenance technology, specifically to a method for monitoring the operating status of wind turbines based on multi-source heterogeneous data. Background Technology

[0002] With the large-scale development of the wind power industry, wind turbines are mostly deployed in harsh environments such as remote mountainous areas and offshore locations. Factors such as strong storms, sandstorms, and salt erosion lead to frequent turbine failures. According to statistics, the cost of wind turbine downtime due to failure accounts for about 20% of the total operation and maintenance cost of a wind farm. Moreover, failures of key components (such as gearbox and generator failures) not only result in high repair costs but also significantly reduce the power generation performance of the turbines, affecting the profitability of the wind farm.

[0003] Operational status monitoring is a core technology for optimizing wind turbine maintenance strategies and reducing operation and maintenance costs, but current monitoring methods face four major technical bottlenecks caused by multi-source heterogeneous SCADA data:

[0004] 1. Data distribution complexity: The multivariate data (such as temperature, vibration, current, and power) collected by the SCADA system come from different sources and have unknown distributions (some variables are periodic and some fluctuate randomly). This does not meet the "multivariate normal distribution" assumption of traditional multivariate statistical process control methods (such as MEWMA control charts), which can easily lead to excessively high false alarm or false alarm rates.

[0005] 2. Grouped Correlation Interference: Due to the structural and functional interrelationships of various components, different collected operating parameters also exhibit correlations, and the strength of these correlations varies depending on the structural location and functional relationship between the components. Components that are structurally closer and functionally more closely related show stronger correlations in their operating states. As the number of sensors increases, the correlation structure becomes more complex. The correlations in wind turbine units exhibit a grouped structure, also known as the clustering effect. When a fault occurs, a group of highly correlated variables often change synchronously, and the clustering effect typically hinders the model's identification of multiple fault variables.

[0006] 3. Sparsity of Fault Signals: Although SCADA systems record a large number of operational variables, when a unit experiences an anomaly, not all components will fail simultaneously. Typically, only a few components malfunction, and only operational variables related to those components will be affected, resulting in abnormal data fluctuations. In high-dimensional SCADA data, if all variables are uniformly modeled and analyzed, the sparse fault characteristics will be masked by the majority of normal signals. Furthermore, the high dimensionality of the variables makes fault identification difficult.

[0007] 4. Insufficient Controlled Samples: The complexity of monitoring models increases with the dimensionality of SCADA data, meaning that training the model requires more controlled samples. However, some newly built wind farms lack sufficient fault-free historical data for model training. For example, some turbines may experience failures early in the wind farm's operation, and there are insufficient controlled samples for these turbines. Furthermore, in current large-scale newly built wind farms, the number of turbines exceeds 50, and the number of SCADA variables collected for each turbine exceeds 100 dimensions. Therefore, rapidly launching monitoring models to achieve accurate whole-machine-level operational status monitoring of newly built wind farms is of practical significance.

[0008] In existing technologies, MEWMA control charts rely on the assumption of normal distribution and have poor adaptability to non-normal SCADA data; while the LASSO method can screen sparse variables, it cannot handle grouped correlations; and although the Random Projected Spatial Rank (RPSR) control chart can handle non-normal data, its fault identification accuracy is only around 80%. Therefore, there is an urgent need for a wind turbine operating status monitoring method that can simultaneously address the challenges of "multi-source heterogeneous distribution, grouped correlations, sparse faults, and insufficient controlled samples". Summary of the Invention

[0009] To address the shortcomings of existing technologies, this invention provides a method for monitoring the operational status of wind turbines based on multi-source heterogeneous data. By converting the unknown distribution of SCADA multi-source heterogeneous data into an approximate multivariate normal distribution, it overcomes the limitations of traditional methods on data distribution. Furthermore, it incorporates an elastic network penalty mechanism to adapt to grouped correlations of variables and sparse fault characteristics, thereby improving fault identification accuracy. This invention aims to enable model startup without requiring a large number of controlled samples, support rapid deployment in newly built wind farms, achieve real-time fault monitoring and precise fault location, and reduce fault response delays.

[0010] To achieve the above-mentioned technical objectives, the present invention provides the following technical solution:

[0011] A method for monitoring the operating status of wind turbine units based on multi-source heterogeneous data, comprising an offline training phase and an online monitoring phase, specifically:

[0012] The offline training phase includes:

[0013] Acquire multi-source heterogeneous SCADA data of wind turbine units, and perform outlier sample removal and standardization processing on the multi-source heterogeneous SCADA data in sequence.

[0014] Spatial rank transformation is performed on the standardized SCADA multi-source heterogeneous data to unify the data distribution and obtain samples that approximate a multivariate normal distribution.

[0015] The samples with an approximate multivariate normal distribution are smoothed; an elastic network penalized likelihood function is constructed based on the smoothed samples; the elastic network penalized likelihood function is solved to obtain the mean vector estimate representing the sparse fault signal in the multi-source heterogeneous data of SCADA.

[0016] Robust optimization of the basic monitoring statistics of wind turbines is performed based on mean vector estimation to obtain robust monitoring statistics; the controlled average running time is set, and the upper control limit of the wind turbine under controlled state is determined through offline simulation.

[0017] The online monitoring phase includes:

[0018] The system collects multi-source heterogeneous data from the SCADA system of the wind turbine in real time and calculates robust monitoring statistics in real time according to the processing flow of the offline training phase. The robust monitoring statistics calculated in real time are compared with the upper control limit determined by the offline simulation. If the robust monitoring statistics exceed the upper control limit, a fault alarm is issued.

[0019] After a fault alarm is triggered, the optimal penalty coefficient is selected, the mean vector estimate is determined based on the optimal penalty coefficient, and the faulty component information is identified and output based on the mean vector estimate.

[0020] Furthermore, the specific steps of sequentially removing outlier samples and standardizing SCADA multi-source heterogeneous data are as follows:

[0021] Remove samples with no data or zero values ​​from the SCADA multi-source heterogeneous data of wind turbine units, which are caused by sensor failure, transmission delay, or downtime maintenance, as well as abnormal fluctuation samples that do not match the fault recording time.

[0022] After removing outlier samples, the SCADA multi-source heterogeneous data is further standardized using Z-score to convert the different operating state variables of the wind turbine into a unified order of magnitude, thus eliminating the influence of dimensions.

[0023] Furthermore, the step of performing spatial rank transformation on the standardized SCADA multi-source heterogeneous data to unify the data distribution and obtain samples with an approximate multivariate normal distribution is as follows:

[0024] Calculate the current time step t before The covariance matrix among the controlled reference samples covariance matrix The sum of the multidimensional operating state variable vectors of wind turbines in each standardized controlled reference sample The mean vector of a controlled reference sample;

[0025] right Perform Cholesky decomposition and take the inverse root to obtain the transformation matrix. ;

[0026] Based on the transformation matrix And the standardized multidimensional running state variable vector in the sampled samples at the current time step t. ,calculate Spatial Rank To characterize the relative size and orientation of the sample;

[0027] For space rank Standardization is performed to obtain samples that approximate a multivariate normal distribution at the current time step t. .

[0028] Furthermore, the process of smoothing the samples that approximate a multivariate normal distribution; constructing an elastic network penalized likelihood function based on the smoothed samples; and solving the elastic network penalized likelihood function to obtain the mean vector estimate representing the sparse fault signal in the multi-source heterogeneous data of SCADA specifically involves:

[0029] An exponentially weighted moving average is used to smooth the sample that approximates a multivariate normal distribution, resulting in a smoothed sample. ,in For smoothing parameters, , These are samples from approximately multivariate normal distributions at time steps t and t-1, respectively, and are set to... ;

[0030] by For input, construct the elastic network penalized likelihood function. Its formula is expressed as:

[0031] ;

[0032] in, For adaptive weight matrix, For balance coefficient, The penalty coefficient is... These are the regression coefficients; For the first The sampling time of the first sampling moment One runtime variable, for The One portion, For the operational variables of the SCADA system;

[0033] remember The Least Angle Regression (LARS) algorithm is used to minimize the penalized likelihood function of the elastic network. Solving for regression coefficients Optimal estimate This leads to the mean vector estimate representing the sparse fault signal. , The zero element in the middle and non-zero elements corresponds to the potential operating state fault variables of the wind turbine.

[0034] Furthermore, the robust optimization of the basic monitoring statistics of the wind turbine based on mean vector estimation is performed to obtain robust monitoring statistics; the upper control limit of the wind turbine under controlled state is determined by setting the controlled average running time and through offline simulation, specifically as follows:

[0035] Based on mean vector estimation Calculate the current time step Basic monitoring statistics The formula is expressed as:

[0036] ;

[0037] in, This is an EWMA weight correction term. For smoothing parameters, for The transpose of the normalized spatial rank vector after EWMA smoothing at time step 1. This is the transpose of the mean drift vector estimated based on the elastic net penalty.

[0038] right Perform robust optimization:

[0039] Construct a sequence of penalty coefficients decreasing from large to small to 0: ;

[0040] definition Indicates the penalty coefficient Optimal estimation of the regression coefficient The set of indices of non-zero elements, denoted as The number of elements in is ;

[0041] Take values ​​one by one from the penalty coefficient sequence, when = At that time, if Make number of elements Add one, then record As a transition point, iterate through all values ​​in the penalty coefficient sequence to construct a system such that... Increase from 1 to The set of transition points ,in, This refers to the number of non-zero regression coefficients. Indicates that from -1 increased to -2 is the transition point; in the absence of Given prior knowledge, the default setting Then the robust monitoring statistics are obtained. :

[0042] ;

[0043] in, , These are the expected value and the variance, respectively. = ;

[0044] Set Controlled Average Run Length (IC-ARL) = Through multiple offline simulations, the method to make IC-ARL= Robust monitoring statistics The value of is the upper control limit.

[0045] Furthermore, after triggering the fault alarm, the optimal penalty coefficient is selected, the mean vector estimate is determined based on the optimal penalty coefficient, and the faulty component information is identified and output based on the mean vector estimate, specifically as follows:

[0046] After a fault alarm is triggered, the optimal penalty coefficient is first selected based on the Risk Inflation Criterion (RIC). Then, by using the elastic net penalized likelihood function and the least angular regression LARS algorithm, the optimal estimate of the regression coefficients under the optimal penalty coefficient is obtained. Finally, based on the adaptive weight matrix Mapping yields mean vector estimate ;

[0047] Each non-zero element corresponds to a fault operation state variable of the wind turbine. The number of fault operation state variables is determined by the number of non-zero elements. Then, based on the mapping relationship between SCADA data and components, the location of the faulty component is determined. Finally, complete faulty component information including fault type, number of faults, and location of the faulty component is output.

[0048] In addition, this application also discloses an electronic device comprising a memory and a processor, wherein:

[0049] Memory is used to store computer programs that can run on a processor;

[0050] The processor is used to execute, as described above, a method for monitoring the operating status of wind turbine generators based on multi-source heterogeneous data when running the computer program.

[0051] A computer-readable storage medium is also disclosed, which stores computer instructions for causing a processor to execute a wind turbine operating status monitoring method based on multi-source heterogeneous data as described above.

[0052] Based on the above technical solution, the present invention has at least the following beneficial effects:

[0053] Strong adaptability to multi-source heterogeneous data: Through spatial rank transformation, SCADA data with unknown distributions such as temperature and vibration are unified into an approximate multivariate normal distribution, which reduces the false alarm rate and solves the limitation of traditional methods on data distribution;

[0054] High fault identification accuracy: The elastic mesh penalized likelihood function combined with L1 and L2 regularization can both screen sparse fault variables and adapt to group correlations, thereby ensuring a high fault identification accuracy while significantly reducing the expected error rate, which is a significant improvement in fault identification performance compared with existing methods.

[0055] Fast startup and excellent real-time performance: The model can be trained with only a small number of fault-free samples, supporting the rapid deployment of new wind farms; Based on recursive optimization of spatial rank and covariance matrix calculation, it ensures that the single sample processing time is shortened and the fault response delay time is fast.

[0056] Significantly reduced operation and maintenance costs: Early warning and precise location of faults reduce unplanned downtime, lower maintenance costs, and help improve the annual revenue of wind farms. Attached Figure Description

[0057] Figure 1 This invention proposes a method for monitoring the operating status of wind turbine units based on multi-source heterogeneous data.

[0058] Figure 2 This is a quantile diagram of the four operating state variables of the wind turbine in an embodiment of the present invention;

[0059] Figure 3 This is a heat map showing the correlation between 12 operating state variables of the wind turbine in this embodiment of the invention;

[0060] Figure 4 The graph shows the results of six online fault monitoring tests conducted based on the method proposed in this invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this invention clearer, the following description is provided in conjunction with the appendix. Figure 1-4 The present invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0062] Although the steps in this invention are arranged by reference numerals, this is not intended to limit the order of the steps. Unless the order of the steps is explicitly stated or the execution of a step requires other steps as a basis, the relative order of the steps can be adjusted. It is understood that the term "and / or" as used herein refers to and covers any and all possible combinations of one or more of the associated listed items.

[0063] like Figure 1 As shown, this invention discloses a method for monitoring the operating status of wind turbine units based on multi-source heterogeneous data, which includes an offline training phase and an online monitoring phase, specifically:

[0064] The offline training phase includes:

[0065] S1. Acquire multi-source heterogeneous SCADA data of the wind turbine, including grid A / B / C phase currents, nacelle X / Y vibration signals and RMS values, active power, reactive power and power factor, rotor and generator speeds, and hydraulic system pressures, etc.

[0066] The experimental environment for the monitoring method in this embodiment was set as follows: CPU: Intel(R) Core(TM) i5-1035G1 CPU @ 1.00GHz, 1.19 GHz. The SCADA multi-source heterogeneous data came from a publicly available dataset of a wind farm. Six different types of faults occurring in different wind turbine units were selected to verify the proposed monitoring method, involving wind turbine units No. 3, 6, 10, 12, 18, and 23. The specific fault descriptions and occurrence times are shown in Table 1.

[0067] Table 1. Detailed information on the 6 fault records

[0068]

[0069] As shown in the table above, the gearbox, generator, hydraulic system, nacelle structure, and converter of wind turbine units are prone to failure and have high failure costs. This embodiment mainly monitors failures from the above-mentioned components and systems. The multi-source state data used involves 12 operating variables in 5 categories: grid A / B / C phase currents, nacelle X / Y vibration signals and RMS values, active power, reactive power and power factor, rotor and generator speeds, and hydraulic system pressure. The 12 variables are numbered V1-V12, and the specific information is shown in Table 2 below. The controlled samples used for offline training are selected from the week before 6 failures (a total of 1008 data points), while the samples used for online monitoring are all samples (144 data points) on the day the failure occurred.

[0070] Table 2 Detailed information on runtime variables

[0071]

[0072] like Figure 2 As shown in the figure (the red line represents the reference number of the theoretical distribution, and the blue line represents the quantile of the actual distribution), this embodiment selected four different types of operating state variables from the above 12 operating state variables: V1 (grid A-phase current), V7 (active power), V10 (wind turbine speed), and V12 (hydraulic system pressure), and plotted their quantile diagrams respectively. It can be seen from the figure that none of the four operating variables satisfy a normal distribution. Therefore, the directly collected SCADA multi-source heterogeneous data cannot meet the prerequisite for the use of traditional multivariate statistical process control methods (the sample follows a multivariate normal distribution). When used to monitor the operating status of wind turbine units, false alarms and missed alarms will occur.

[0073] For example Figure 3 As shown, the correlation heatmap of the 12 operating state variables shows stronger correlations among data variables of the same type. For example, there are grouped correlation structures between V1-V3, V4-V6, V7-V9, and V10 and V11. Furthermore, variables that are more similar in function and mechanism also have stronger correlations. For example, the wind turbine speed and generator speed affect how much electricity the unit can generate, i.e., they are related to active power. Active power is positively correlated with the three-phase current of the generator. Therefore, V1-V3 also have grouped correlations with V7, V8, and V10, V11 (i.e., grouping effect, which can mask coefficient fault characteristics and affect monitoring accuracy). Therefore, to address the complex distribution properties and grouping effect of multi-source heterogeneous data, the state monitoring method proposed in this invention performs the following steps during the offline training phase.

[0074] After collecting multi-source heterogeneous SCADA data, outlier removal and standardization processing should be performed sequentially. In this preferred embodiment, the outlier removal and standardization processing specifically involves:

[0075] Remove samples with no data or zero values ​​from the SCADA multi-source heterogeneous data of wind turbine units, which are caused by sensor failure, transmission delay, or downtime maintenance, as well as abnormal fluctuation samples that do not match the fault recording time (such as samples with a current surge of 10 times the mean).

[0076] After removing outlier samples, the SCADA multi-source heterogeneous data is further standardized using Z-score to convert the different operating state variables of the wind turbine into a unified order of magnitude, thus eliminating the influence of dimensions.

[0077] S2. Perform spatial rank transformation on the standardized SCADA multi-source heterogeneous data to unify the data distribution and obtain samples that approximate a multivariate normal distribution.

[0078] In a preferred embodiment, step S2 specifically comprises:

[0079] Calculate the current time step t before The covariance matrix among the controlled reference samples covariance matrix The sum of the multidimensional operating state variable vectors of the wind turbine in each standardized controlled reference sample (the so-called controlled reference sample is the SCADA multivariate heterogeneous data sample collected by the wind turbine under normal operating conditions) The mean vector of the controlled reference samples is expressed by the formula:

[0080] ;

[0081] in, This represents a multidimensional operating state variable vector of a wind turbine. express The mean vector of a controlled reference sample;

[0082] right Perform Cholesky decomposition and take the inverse root to obtain the transformation matrix. (satisfy );

[0083] Based on the transformation matrix And the standardized multidimensional running state variable vector in the sampled samples at the current time step t. ,calculate Spatial Rank This characterizes the relative size and orientation of a sample; where relative size reflects the distance between the current sample and the center of the reference sample, and orientation reflects the specific direction of deviation of the current sample; the formula for calculating the spatial rank is:

[0084] ;

[0085] in, For spatial symbolic functions;

[0086] For space rank Standardization is performed to obtain samples that approximate a multivariate normal distribution at the current time step t. The formula is expressed as:

[0087] ;

[0088] in, The recursive estimation formula based on historical samples is:

[0089] , To ensure The covariance matrix is ​​a unit diagonal matrix.

[0090] This application unifies the unknown distribution of multi-source heterogeneous SCADA data into an approximate multivariate normal distribution, overcoming the limitations of traditional methods on data distribution and helping to reduce the false alarm rate of fault monitoring.

[0091] S3. Smooth the samples that are approximately multivariately normally distributed; construct an elastic network penalized likelihood function based on the smoothed samples. The elastic network penalized likelihood function combines L1 and L2 regularization, which can both screen sparse fault variables and adapt to group correlations; solve the elastic network penalized likelihood function to obtain the mean vector estimate that accurately represents the sparse fault signals in SCADA multi-source heterogeneous data.

[0092] In a preferred embodiment, step S3 specifically comprises:

[0093] An exponentially weighted moving average is used to smooth the sample that approximates a multivariate normal distribution, resulting in a smoothed sample. ,in For smoothing parameters, , These are samples from approximately multivariate normal distributions at time steps t and t-1, respectively, and are set to... ;

[0094] by For input, construct the elastic network penalized likelihood function. Its formula is expressed as:

[0095] ;

[0096] in, For adaptive weight matrix, For balance coefficient, The penalty coefficient is... For regression coefficients (set in this embodiment) and Consistent dimensions The non-zero elements in the equation correspond to the fault variables and their degree of offset. For the first The sampling time of the first sampling moment One runtime variable, for The One portion, For the operational variables of the SCADA system;

[0097] remember The Least Angle Regression (LARS) algorithm is used to minimize the penalized likelihood function of the elastic network. Solving for regression coefficients Optimal estimate This leads to the mean vector estimate representing the sparse fault signal. , The zero element in the middle and non-zero elements corresponds to the potential operating state fault variables of the wind turbine.

[0098] S4. Based on mean vector estimation, robust optimization of the basic monitoring statistics of wind turbine units is performed to obtain robust monitoring statistics; the controlled average running time is set, and the upper control limit of wind turbine units under controlled state is determined through offline simulation.

[0099] In a preferred embodiment, step S4 specifically includes:

[0100] Based on mean vector estimation Calculate the current time step Basic monitoring statistics The formula is expressed as:

[0101] ;

[0102] in, This is an EWMA weight correction term; for The transpose of the normalized spatial rank vector after EWMA smoothing at time step 1. This is the transpose of the mean drift vector estimated based on the elastic net penalty.

[0103] For each penalty coefficient corresponding ,when When the value of is too large, it means a significant penalty to the mean vector estimation, causing the regression coefficients of each dimension to tend to zero, making fault identification difficult and increasing the risk of unplanned downtime and maintenance costs. Therefore, this application determines the regression coefficient set based on the number of non-zero elements in the regression coefficients and generates a corresponding mean vector estimate to correct the error. That is to The following robustness optimizations were performed:

[0104] Construct a sequence of penalty coefficients decreasing from large to small to 0: ;

[0105] definition Indicates the penalty coefficient Optimal estimation of the regression coefficient The set of indices of non-zero elements, denoted as The number of elements in is ;

[0106] Take values ​​one by one from the penalty coefficient sequence, when = At that time, if Make number of elements Add one, then record As a transition point, iterate through all values ​​in the penalty coefficient sequence to construct a system such that... Increase from 1 to The set of transition points ,in, Indicates that from -1 increased to The transition point of -2, that is, dividing the original penalty coefficient sequence into segments:

[0107] ;

[0108] about To obtain the solution, this embodiment provides an easy-to-implement solution: set a sufficiently small initial value. and ,initial Perform iterative calculations ,when At that time, the set of transition points is expanded to Repeat the iteration until To obtain the final Furthermore, in the absence of Given prior knowledge, this application defaults to the following settings. Then the robust monitoring statistics are obtained. :

[0109] ;

[0110] in, , These are the expected value and the variance, respectively. = ;

[0111] Set Controlled Average Run Length (IC-ARL) = (In this embodiment, the value is taken as 200). Through 10,000 offline simulations, the value of IC-ARL= is found. Robust monitoring statistics The value of is the upper control limit. In this application, the robust monitoring statistic, by introducing a penalty coefficient sequence and a transition point set, can effectively address non-normal distribution and group effects, significantly improving the monitoring sensitivity and accuracy of sparse fault signals.

[0112] The offline training is now complete. Following the above process, this invention constructs a multivariate nonparametric statistical process control chart (MVNP control chart), which can be used to achieve real-time online monitoring of the operating state variables of wind turbine units. The online monitoring phase includes:

[0113] S5. Real-time acquisition of SCADA multi-source heterogeneous data from wind turbine units, and real-time calculation of robust monitoring statistics according to the processing flow of the offline training phase; compare the real-time calculated robust monitoring statistics with the upper control limit determined by offline simulation, and issue a fault alarm if the robust monitoring statistics exceed the upper control limit.

[0114] S6. After triggering the fault alarm, the optimal penalty coefficient is selected, the mean vector estimate is determined based on the optimal penalty coefficient, and the faulty component information is identified and output based on the mean vector estimate. In a preferred embodiment, step S6 specifically involves:

[0115] After a fault alarm is triggered, the optimal penalty coefficient is first selected based on the Risk Inflation Criterion (RIC). Its formula is expressed as:

[0116] ;in, for and Poor transpose The time when the fault occurred. The sample is smoothed out at the moment the fault occurred. To characterize the mean vector estimation of sparse fault signals, yes The number of non-zero elements in the neutron. This is a penalty item according to RIC criteria;

[0117] Then, by using the elastic net penalized likelihood function and the least angle regression (LARS) algorithm, the optimal estimate of the regression coefficients under the optimal penalty coefficient is obtained. Finally, based on the adaptive weight matrix Mapping yields mean vector estimate ;

[0118] Each non-zero element corresponds to a fault operation state variable of the wind turbine. The number of fault operation state variables is determined by the number of non-zero elements. Then, based on the mapping relationship between SCADA data and components, the location of the faulty component is determined. Finally, complete faulty component information, including fault type, number of faults, and location of the faulty component, is output. This concludes the complete process of the condition monitoring method proposed in this invention.

[0119] In addition, this application also discloses an electronic device comprising a memory and a processor, wherein:

[0120] Memory is used to store computer programs that can run on a processor;

[0121] The processor is used to execute, as described above, a method for monitoring the operating status of wind turbine generators based on multi-source heterogeneous data when running the computer program.

[0122] This application also discloses a computer-readable storage medium storing computer instructions, which are used to cause a processor to execute a wind turbine operating status monitoring method based on multi-source heterogeneous data as described above.

[0123] The following embodiment presents an analysis and comparison of the results of the proposed method of this invention with existing methods in the condition monitoring of wind turbine units:

[0124] First, the method proposed in this invention was used to monitor the six faults in Table 1, and the results are as follows: Figure 4 As shown: the red area represents the time period in which the fault occurred, and the blue solid line represents the robust monitoring statistics. The red dashed line represents the upper control limit (UCL). It can be seen that when monitoring wind turbines using the MVNP control chart constructed based on this invention, the statistical quantities during the turbine fault phase and normal operation phase are clearly distinguished. When the turbine is operating normally, the monitored statistical quantities... All values ​​are below the control limits, and the monitoring scheme does not issue false alarms at this time. However, when the unit operates to the fault stage, the monitoring statistics... The temperature rises rapidly until it exceeds the control limit, triggering an alarm from the unit. Once the unit returns to normal operation, meaning the fault is repaired, the monitoring statistics return to normal. It has fallen below the control limit.

[0125] The monitoring results from the six fault records show that although the types, durations, and forms of the faults are different, the proposed monitoring method can respond quickly, issue an alarm at the first moment the fault occurs, and quickly restore the normal state after the fault is reset. It has high detection accuracy and low false alarm rate.

[0126] To verify the performance difference between the proposed method and current mainstream monitoring methods, this example first compares the fault detection performance, selecting the following existing methods:

[0127] 1. Random projection spatial rank (RPSR) control chart;

[0128] 2. Distribution-free (DF) EWMA control chart;

[0129] 3. Multivariate sign (MS) EWMA control chart;

[0130] 4. MEWMA control chart;

[0131] 5. Spatial Rank (SR) EWMA Control Chart;

[0132] 6. Self-starting (SS) EWMA control chart.

[0133] As a comparative model, the numerical experiment was set up as follows: number of fault-free samples Data dimensions Location of the fault Table 3 shows and OC-ARL (Spatiotemporal Average Runtime, with smaller values ​​indicating higher monitoring sensitivity) results were obtained under three distributions: multivariate normal, multivariate t-distribution, and multivariate Gamma distribution. To ensure robustness, each monitoring result was calculated based on the average of ten thousand simulations. The UCL of all monitoring methods was determined according to... , The sample distribution was adjusted so that IC-ARL=200. The results are shown in Table 3:

[0134] Table 3. Comparison of OC-ARL using different methods at p=10

[0135]

[0136] in The drift magnitude is represented by the number of observations. Under a multivariate normal distribution, MVNP performs well in monitoring large-scale drifts. Compared to MEWMA and SSEWMA, which are based on raw observations, other sign- or rank-based methods lose some distance information, thus their monitoring performance is relatively poor under a multivariate normal distribution. For non-normal distributions, MEWMA's performance deteriorates significantly, while SREN shows a monitoring advantage. For multivariate... And multivariate Gamma distribution, SREN control plot in It performs best when the data is in the range of rank. Specifically, because MVNP uses spatial rank instead of spatial symbols, it contains more distance information compared to MSEWMA, thus responding faster when detecting large-scale drifts. Furthermore, due to the addition of elastic net penalty, MVNP's adaptability to group effects when strongly correlated variables drift simultaneously further improves its monitoring performance. Therefore, its monitoring capability for medium to large displacements is superior to other rank-based comparative methods. However, for smaller drifts, the variable selection mechanism with L2 penalty may compress all variables to near-zero levels, leading to a decrease in monitoring performance. Unlike spatial rank methods that rely on symmetry assumptions, when... When the size is smaller, the monitoring performance of DFEWMA is significantly improved. Therefore, in Next, when and At that time, DFEWMA outperformed all other methods. From the above performance comparison, it can be seen that, with a fixed drift direction, for non-normal samples, especially for medium and large drifts, the monitoring performance of this method has certain advantages over other monitoring methods. In the condition monitoring of wind turbine units, faults are mostly characterized by significant mean drift; therefore, the method proposed in this invention can effectively monitor unit faults.

[0137] In practice, after monitoring abnormal signals from wind turbines, it is necessary to further identify specific faulty components in order to arrange targeted maintenance strategies and spare parts replacement plans. Therefore, this embodiment further compares the performance of fault variable identification: the method proposed in this invention is compared with common fault identification methods under both normal and non-normal distributions. Under a multivariate normal distribution, this example selects the following three fault variable identification methods:

[0138] 1. Step-down filtering;

[0139] 2. Fault variable diagnosis based on LASSO;

[0140] 3. Adaptive step-down (ASD) filtering;

[0141] The diagnostic process is evaluated using two metrics: correctness ratio (CR) and expected error rates (EER). CR can be expressed as:

[0142] ;

[0143] in, For the number of times the identification is repeated, This is an indicator function that equals 1 when all faulty components are correctly identified.

[0144] The formula for calculating EER is:

[0145] ;

[0146] in, This indicates the number of unidentified faulty components. A larger CR and a smaller EER indicate better performance of the fault variable identification method.

[0147] Assume the controlled operation of the wind turbine follows a multivariate normal distribution: mean covariance matrix Among them, the independent variance Covariance ,set up When a fault occurs during operation, the mean drift is set to... ,in , , This indicates that variable 1 has drifted. Other variables do not drift, among which This indicates the drift direction, i.e., the location of the fault variable.

[0148] In the above scenario, Table 4 shows the performance of the proposed method in fault variable identification compared with several existing methods:

[0149] Table 4. Recognition performance of different methods under normal distribution.

[0150]

[0151] Table 4 shows that the MVNP control chart constructed by our method outperforms other methods in overall identification performance, and this advantage is evident in various drift scenarios. When there is only one faulty component, MVNP exhibits high accuracy, indicating its ability to more accurately diagnose sparse faults. As the number of faulty components increases, the identification performance of all methods decreases. However, compared to other methods, MVNP shows the slowest decline. Therefore, it can also identify the simultaneous shift of multiple variables. Notably, MVNP's performance further improves when the directions of fault components are close, meaning that MVNP can identify highly correlated fault variables. This is a significant advantage over LASSO, confirming that resilient networks can handle grouping effects more effectively than LASSO.

[0152] In summary, this invention proposes a wind turbine operation status monitoring method that can simultaneously address multi-source heterogeneous distribution, group correlation, sparse faults, and insufficient controlled samples. Compared with existing methods, it has a better recognition accuracy, a lower false alarm rate, and can quickly and accurately monitor multiple types of faults, thereby reducing operation and maintenance costs.

[0153] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0154] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus or device (such as a computer-based system, a processor-included system or other system that can fetch and execute instructions from, an instruction execution system, apparatus or device).

[0155] The above embodiments provide a detailed description of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for monitoring the operating status of wind turbine units based on multi-source heterogeneous data, characterized in that, It includes an offline training phase and an online monitoring phase, specifically: The offline training phase includes: Acquire multi-source heterogeneous SCADA data of wind turbine units, and perform outlier sample removal and standardization processing on the multi-source heterogeneous SCADA data in sequence. Spatial rank transformation is performed on the standardized SCADA multi-source heterogeneous data to unify the data distribution and obtain samples that approximate a multivariate normal distribution. The samples with an approximate multivariate normal distribution are smoothed; an elastic network penalized likelihood function is constructed based on the smoothed samples; the elastic network penalized likelihood function is solved to obtain the mean vector estimate representing the sparse fault signal in the multi-source heterogeneous data of SCADA. Robust optimization of the basic monitoring statistics of wind turbines is performed based on mean vector estimation to obtain robust monitoring statistics; the controlled average runtime is set, and the upper control limit of the wind turbine under controlled state is determined through offline simulation, specifically: Based on mean vector estimation Calculate the current time step Basic monitoring statistics The formula is expressed as: ； in, This is an EWMA weight correction term. For smoothing parameters, for The transpose of the normalized spatial rank vector after EWMA smoothing at time step 1. This is the transpose of the mean drift vector estimated based on the elastic net penalty. right Perform robust optimization: Construct a sequence of penalty coefficients decreasing from large to small to 0: ; definition Indicates the penalty coefficient Optimal estimation of the regression coefficient The set of indices of non-zero elements, denoted as The number of elements in is ; Take values ​​one by one from the penalty coefficient sequence, when = At that time, if Make number of elements Add one, then record As a transition point, iterate through all values ​​in the penalty coefficient sequence to construct a system such that... Increase from 1 to The set of transition points ,in, This refers to the number of non-zero regression coefficients. Indicates that from Increase to The transition point; in the absence of Given prior knowledge, the default setting Then the robust monitoring statistics are obtained. : ； in, , These are the expected value and the variance, respectively. = ; Set Controlled Average Run Length (IC-ARL) = Through multiple offline simulations, the method to make IC-ARL= Robust monitoring statistics The value of is the upper control limit; The online monitoring phase includes: The system collects multi-source heterogeneous data from the SCADA system of the wind turbine in real time and calculates robust monitoring statistics in real time according to the processing flow of the offline training phase. The robust monitoring statistics calculated in real time are compared with the upper control limit determined by the offline simulation. If the robust monitoring statistics exceed the upper control limit, a fault alarm is issued. After a fault alarm is triggered, the optimal penalty coefficient is selected, the mean vector estimate is determined based on the optimal penalty coefficient, and the faulty component information is identified and output based on the mean vector estimate.

2. The method for monitoring the operating status of wind turbine units based on multi-source heterogeneous data according to claim 1, characterized in that, The specific steps for sequentially removing outliers and standardizing SCADA multi-source heterogeneous data are as follows: Remove samples with no data or zero values ​​from the SCADA multi-source heterogeneous data of wind turbine units, which are caused by sensor failure, transmission delay, or downtime maintenance, as well as abnormal fluctuation samples that do not match the fault recording time. After removing outlier samples, the SCADA multi-source heterogeneous data is further standardized using Z-score to convert the different operating state variables of the wind turbine into a unified order of magnitude, thus eliminating the influence of dimensions.

3. The method for monitoring the operating status of wind turbine units based on multi-source heterogeneous data according to claim 1, characterized in that, The specific steps for performing spatial rank transformation on the standardized SCADA multi-source heterogeneous data to unify the data distribution and obtain samples with an approximate multivariate normal distribution are as follows: Calculate the current time step t before The covariance matrix among the controlled reference samples covariance matrix The sum of the multidimensional operating state variable vectors of wind turbines in each standardized controlled reference sample The mean vector of a controlled reference sample; right Perform Cholesky decomposition and take the inverse root to obtain the transformation matrix. ; Based on the transformation matrix And the standardized multidimensional running state variable vector in the sampled samples at the current time step t. ,calculate Spatial Rank To characterize the relative size and orientation of the sample; For space rank Standardization is performed to obtain samples that approximate a multivariate normal distribution at the current time step t. .

4. The method for monitoring the operating status of wind turbine units based on multi-source heterogeneous data according to claim 1, characterized in that, The process involves smoothing the samples that approximate a multivariate normal distribution; constructing an elastic network penalized likelihood function based on the smoothed samples; and solving the elastic network penalized likelihood function to obtain the mean vector estimate representing the sparse fault signal in the multi-source heterogeneous data of SCADA. Specifically, this involves: An exponentially weighted moving average is used to smooth the sample that approximates a multivariate normal distribution, resulting in a smoothed sample. ,in For smoothing parameters, , These are samples from approximately multivariate normal distributions at time steps t and t-1, respectively, and are set to... ; by For input, construct the elastic network penalized likelihood function. Its formula is expressed as: ； in, For adaptive weight matrix, For balance coefficient, The penalty coefficient is... These are the regression coefficients; For the first The sampling time of the first sampling moment One runtime variable, for The One portion, For the operational variables of the SCADA system; remember The Least Angle Regression (LARS) algorithm is used to minimize the penalized likelihood function of the elastic network. Solving for regression coefficients Optimal estimate This leads to the mean vector estimate representing the sparse fault signal. , The zero element in the middle and non-zero elements corresponds to the potential operating state fault variables of the wind turbine.

5. The method for monitoring the operating status of wind turbine units based on multi-source heterogeneous data according to claim 1, characterized in that, After triggering the fault alarm, the optimal penalty coefficient is selected, the mean vector estimate is determined based on the optimal penalty coefficient, and the faulty component information is identified and output based on the mean vector estimate. Specifically: After a fault alarm is triggered, the optimal penalty coefficient is first selected based on the Risk Inflation Criterion (RIC). Then, by using the elastic net penalized likelihood function and the least angular regression LARS algorithm, the optimal estimate of the regression coefficients under the optimal penalty coefficient is obtained. Finally, based on the adaptive weight matrix Mapping yields mean vector estimate ; Each non-zero element corresponds to a fault operation state variable of the wind turbine. The number of fault operation state variables is determined by the number of non-zero elements. Then, based on the mapping relationship between SCADA data and components, the location of the faulty component is determined. Finally, complete faulty component information including fault type, number of faults, and location of the faulty component is output.

6. An electronic device, characterized in that, The electronic device includes a memory and a processor, wherein: Memory is used to store computer programs that can run on a processor; A processor is configured to execute, when running the computer program, a wind turbine operating status monitoring method based on multi-source heterogeneous data as described in any one of claims 1-5.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that cause a processor to execute a method for monitoring the operating status of a wind turbine based on multi-source heterogeneous data as described in any one of claims 1-5.

Images