Method and system for detecting abnormal operation data of a wind turbine generator and medium

Abstract

The application provides a wind turbine operation data anomaly detection method, system and medium. The method comprises the following steps: removing abnormal data and non-steady-state data from historical operation data of a target wind turbine; dividing the historical operation data into working conditions through equal-interval classification, clustering data samples under each working condition according to sample similarity through a Gaussian mixture model, and screening reference samples according to a preset screening standard; constructing a reference model of the target wind turbine under variable working conditions based on a conditional variational autoencoder, and training the reference model through the reference samples; calculating a reconstruction probability through the trained reference model, obtaining real-time operation data of the target wind turbine, and positioning abnormal operation data according to the reconstruction probability and a deviation degree of real-time operation parameters. The method can establish an accurate reference model under variable working conditions, improves the accuracy of operation state evaluation of the unit under variable working conditions, and accurately positions abnormal parameters.

Description

Technical Field

[0001] This application relates to the field of wind power generation technology, and in particular to a method, system and medium for detecting abnormal operating data of wind turbine generators. Background Technology

[0002] As the adoption rate of wind power continues to increase, greater attention is being paid to maintaining the safe and stable operation of wind turbines. Monitoring and assessing the health status of wind turbines is a crucial foundation for ensuring their safe and stable production. Among these measures, anomaly detection in wind turbine operating data assesses whether the turbine's operating condition is abnormal under current conditions. This is a prerequisite for conducting health maintenance work on wind turbines, and the operating data of the turbine when it is in a healthy operating state forms the basis for predicting wind turbine power generation. Currently, anomaly detection is generally achieved by monitoring and analyzing real-time operating data of the wind turbine, comparing the current data with baseline data, and thus identifying anomalies. Therefore, the core of anomaly detection in wind turbine operating conditions is the establishment of a baseline model.

[0003] In related technologies, the usual approach is to first conduct performance tests on the wind turbine, and then build a benchmark model for evaluating the turbine's operating status based on the test results and relevant mathematical models. However, due to the inter-coupling of wind turbine operating parameters and the influence of external boundary conditions, they exhibit high-dimensional, dynamic, and nonlinear characteristics. Furthermore, as wind turbine equipment ages and external conditions change, the actual benchmark state of the turbine often deviates significantly from the design values. Therefore, the methods described above cannot meet the need for benchmark modeling of wind turbines under changing operating conditions. The established models are not applicable to variable operating condition scenarios in practical applications, resulting in low accuracy in detecting anomalies in the wind turbine's operating status and an inability to promptly identify any abnormalities. Summary of the Invention

[0004] This application aims to at least partially address one of the technical problems in the related art.

[0005] Therefore, the first objective of this application is to propose a method for detecting anomalies in the operating data of wind turbine units. This method uses a variable autoencoder model to reconstruct data under varying operating conditions, establishes an anomaly detection model that integrates multiple feature parameters, and realizes real-time dynamic assessment of the health status of wind turbine units under varying operating conditions. This allows for the establishment of an accurate benchmark model under varying operating conditions, improves the accuracy of the assessment of the unit's operating status under varying operating conditions, and can accurately locate abnormal parameters.

[0006] The second objective of this application is to propose a system for detecting abnormal operating data of wind turbine generators;

[0007] The third objective of this application is to provide a non-transitory computer-readable storage medium.

[0008] To achieve the above objectives, a first aspect of this application provides a method for detecting abnormal operating data of a wind turbine generator, the method comprising the following steps:

[0009] Obtain historical operating data of the target wind turbine to be tested, and remove abnormal data and non-steady-state data that do not conform to the output characteristics of the wind turbine from the historical operating data;

[0010] Select operating condition characteristic parameters, divide the historical operating data into operating conditions through equal interval classification, and cluster the data samples under each operating condition according to sample similarity using Gaussian Mixture Model (GMM). Select benchmark samples from all clusters under each operating condition according to preset screening criteria.

[0011] The operating status of wind turbines under varying operating conditions is analyzed. A baseline model of the target wind turbine under varying operating conditions is constructed based on a conditional variational autoencoder. The conditional variational autoencoder is then trained using the baseline samples to obtain the trained baseline model.

[0012] The reconstruction probability is calculated using the trained benchmark model as a feature index for anomaly detection. Real-time operating data of the target wind turbine is obtained, and abnormal operating data is located based on the reconstruction probability and the deviation between the actual value and the reconstructed value of the real-time operating data.

[0013] Optionally, in one embodiment of this application, removing non-steady-state data from the historical operating data includes: expressing the output power of the target wind turbine as an expression that includes the rate of change of power; calculating the difference in output power between two adjacent moments; estimating the rate of change of power using the mean of the sample statistics within the time window; determining the confidence interval in which the true value of the rate of change of power lies using interval estimation; and determining whether the target wind turbine is in a non-steady-state condition within the time window based on whether the confidence interval does not include zero. If so, the non-steady-state data is removed.

[0014] Optionally, in one embodiment of this application, the selected operating condition characteristic parameters include: wind speed, wind direction, torque, rotational speed, ambient temperature, and load. The step of classifying the historical operating data into operating conditions through equal-interval classification includes: determining the maximum and minimum values ​​of each operating condition characteristic parameter within its own range of variation, and obtaining a preset operating condition classification interval for each operating condition characteristic parameter; the equal-interval operating condition classification is performed using the following formula:

[0015] C i =(S i ,S i +ΔS)∩(Di D i +ΔD)∩(M i M i +ΔM)∩(N i N i +ΔN)∩(T i ,T i +ΔT)∩(P i ,P i +ΔP)

[0016] stS i <<S min ,S i +ΔS>>S max D i <<D min D i +ΔD>>D max M i <<M min M i +ΔM>>M max N i <<N min N i +ΔN>>N max ,T i <<T min ,T i +ΔT>>T max ,P i <<P min ,P i +ΔP>>P max

[0017] Where S is wind speed, D is wind direction, M is torque, N is rotational speed, T is ambient temperature, and P is load; invalid working conditions are removed from all the divided working conditions until the number of remaining valid working conditions is greater than the preset proportion of the total number of theoretical working conditions.

[0018] Optionally, in one embodiment of this application, the data samples under each operating condition are clustered based on sample similarity using a Gaussian Mixture Model (GMM), including: estimating the parameters of the GMM using the Expectation-Maximization (EM) algorithm; determining the number of sub-models of the GMM using the Akaike Information Criterion (AIC) to cluster the state categories of the target wind turbine; and selecting benchmark samples from all clusters under each operating condition according to preset screening criteria, including: comparing the data samples under each cluster and selecting the target data group with the highest average operating efficiency as the benchmark sample; and removing sample points with a likelihood much smaller than the average likelihood using the GMM.

[0019] Optionally, in one embodiment of this application, calculating the reconstruction probability using the trained benchmark model includes: obtaining test samples; obtaining the first and second parameters in the Gaussian distribution of each latent variable using the encoder of the conditional variational autoencoder, and sampling a preset number of sample points for each latent variable; calculating the third and fourth parameters in the likelihood distribution of the reconstruction variable corresponding to each latent variable using the decoder of the conditional variational autoencoder; and calculating the average value of the log-likelihood of the test samples under the latent variable condition based on the third and fourth parameters.

[0020] Optionally, in one embodiment of this application, locating abnormal running data based on the reconstruction probability and the deviation between the actual value and the reconstructed value of the real-time running data includes: estimating the reconstruction probability distribution of the training samples using kernel density estimation (KDE), and using the lower limit of the confidence interval corresponding to the confidence level equal to a preset value as the threshold for anomaly detection; determining whether the reconstruction probability of the real-time running data is abnormal based on the anomaly detection threshold; and calculating the deviation of each parameter in the real-time running data using the following formula when an abnormal reconstruction probability is detected:

[0021]

[0022] Among them, v k x represents the standardized coefficient. k This represents the actual value of parameter k. This represents the reconstructed value of parameter k; by comparing the deviation of each parameter, the parameter that caused the anomaly is determined.

[0023] To achieve the above objectives, a second aspect of this application also proposes a wind turbine operating data anomaly detection system, comprising the following modules:

[0024] The acquisition module is used to acquire the historical operating data of the target wind turbine to be detected, and to remove abnormal data and non-steady-state data that do not conform to the output characteristics of the wind turbine from the historical operating data;

[0025] The filtering module is used to select operating condition characteristic parameters, divide the historical operating data into operating conditions through equal interval classification, and cluster the data samples under each operating condition according to sample similarity using Gaussian Mixture Model (GMM). Based on the preset filtering criteria, the module selects benchmark samples from all clusters under each operating condition.

[0026] The module is used to analyze the operating status of wind turbines under varying operating conditions, construct a benchmark model of the target wind turbine under varying operating conditions based on a conditional variational autoencoder, and train the conditional variational autoencoder using the benchmark samples to obtain the trained benchmark model.

[0027] The detection module is used to calculate the reconstruction probability as a feature index for anomaly detection using the trained benchmark model, and to obtain the real-time operating data of the target wind turbine. Based on the reconstruction probability and the deviation between the actual value and the reconstructed value of the real-time operating data, the abnormal operating data is located.

[0028] Optionally, in one embodiment of this application, the acquisition module is specifically used for: expressing the output power of the target wind turbine as an expression that includes the rate of change of power; calculating the difference in output power between two adjacent moments; estimating the rate of change of power using the mean of the sample statistics within the time window; determining the confidence interval in which the true value of the rate of change of power lies through interval estimation; and determining whether the target wind turbine is in a non-steady-state condition within the time window based on whether the confidence interval does not include zero, and if so, removing the non-steady-state data.

[0029] Optionally, in one embodiment of this application, the selected operating condition characteristic parameters include: wind speed, wind direction, torque, rotational speed, ambient temperature and load. The filtering module is specifically used to: determine the maximum and minimum values ​​of each operating condition characteristic parameter within its own range of variation, and obtain the preset operating condition division interval for each operating condition characteristic parameter.

[0030] The following formula is used to divide the operating conditions into equal intervals:

[0031] C i =(S i ,S i +ΔS)∩(D i D i +ΔD)∩(M i M i +ΔM)∩(N i N i +ΔN)∩(T i ,T i +ΔT)∩(P i ,P i +ΔP)

[0032] stS i <<S min ,S i +ΔS>>S max D i <<D min D i +ΔD>>D max M i <<M min M i +ΔM>>M max N i<<N min N i +ΔN>>N max ,T i <<T min ,T i +ΔT>>T max ,P i <<P min ,P i +ΔP>>P max

[0033] Where S is wind speed, D is wind direction, M is torque, N is rotational speed, T is ambient temperature, and P is load; invalid working conditions are removed from all the divided working conditions until the number of remaining valid working conditions is greater than the preset proportion of the total number of theoretical working conditions.

[0034] To implement the above embodiments, a third aspect of this application also proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the wind turbine operation data anomaly detection method in the above embodiments.

[0035] The technical solution provided by the embodiments of this application brings at least the following beneficial effects: This application first initiates a data cleaning process, establishing steady-state discrimination, operating condition division, and benchmark screening models. Invalid abnormal data, non-steady-state data, and a small number of minor abnormal data points mixed in with the historical data of wind turbine units are removed. The cleaned data is used for training the subsequent anomaly detection model, which can improve the accuracy of the trained anomaly detection model in detecting abnormal data. Then, considering the dynamic complexity of wind turbine unit state information, the correlation between multiple parameters in the data is analyzed in a data-driven manner. A variable autoencoder model is used to reconstruct the data under varying operating conditions, establishing a multi-feature parameter fusion anomaly detection model to achieve real-time dynamic assessment of the health status of wind turbine units under varying operating conditions. Therefore, this method can establish an accurate benchmark model in scenarios where operating conditions change and deviate from the design conditions. This facilitates timely and accurate detection of abnormal parameters under varying operating conditions, improves the accuracy of assessing the operating status of wind turbine units under varying operating conditions, and can accurately locate specific abnormal parameters, allowing staff to obtain more detailed anomaly information and facilitating timely troubleshooting.

[0036] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0037] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0038] Figure 1 This is a flowchart of a method for detecting abnormal operating data of a wind turbine generator according to an embodiment of this application;

[0039] Figure 2 This is a flowchart illustrating a method for steady-state screening of historical operating data of wind turbines, as proposed in an embodiment of this application.

[0040] Figure 3 A flowchart illustrating a method for screening benchmark samples of wind turbine generators according to an embodiment of this application;

[0041] Figure 4 This is a schematic diagram illustrating the relationship between characteristic indicators of a wind turbine under varying operating conditions, as proposed in an embodiment of this application.

[0042] Figure 5 This is a schematic diagram of the structure of a CVAE model proposed in an embodiment of this application;

[0043] Figure 6 This is a flowchart illustrating a specific method for detecting abnormal operating data of a wind turbine generator as proposed in an embodiment of this application;

[0044] Figure 7 This is a schematic diagram of the structure of a wind turbine operation data anomaly detection system proposed in an embodiment of this application. Detailed Implementation

[0045] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0046] The following describes in detail, with reference to the accompanying drawings, a method and system for detecting abnormal operating data of a wind turbine generator proposed in an embodiment of the present invention.

[0047] Figure 1 This is a flowchart of a method for detecting abnormal operating data of a wind turbine generator proposed in an embodiment of this application, as shown below. Figure 1 As shown, the method includes the following steps:

[0048] Step S101: Obtain the historical operating data of the target wind turbine to be tested, and remove abnormal data and non-steady-state data that do not conform to the output characteristics of the wind turbine from the historical operating data.

[0049] Abnormal data refers to numerous anomalies in the measured operating data of wind turbines that do not conform to the normal output characteristics of the wind turbines, caused by factors such as wind curtailment, communication equipment failure, extreme weather, blade contamination, and wind speed sensor malfunction. Non-steady-state data refers to data generated when the wind turbines are operating under unpredictable and unsteady conditions.

[0050] Specifically, the first step is to acquire historical operating data of the target wind turbine for which anomaly detection needs to be performed through various methods. For example, a Supervisory Control and Data Acquisition (SCADA) system can be pre-set in the wind turbine system to monitor various devices of the turbine, and then the operating data of the turbine at different times can be acquired from the SCADA system.

[0051] Furthermore, abnormal data that does not conform to the output characteristics of the wind turbine is removed from the acquired historical operating data. One possible approach is to first identify the abnormal data to be removed. Specifically, industry standards for determining whether a wind turbine is operating normally can be referenced. Operating data generated under the following five scenarios are considered abnormal data that do not conform to the wind turbine's output characteristics: data where external conditions other than wind speed exceed the wind turbine's operating range; data during periods when a fault prevents the turbine from operating; data when the wind turbine is manually shut down or under testing or maintenance; data during periods of test equipment failure or performance degradation, such as blade icing and contamination; and data where the wind direction exceeds the specified measurement sector. Then, all operating data generated within the time periods corresponding to these five scenarios are removed from the historical operating data.

[0052] Therefore, this application eliminates abnormal data before training the detection model to detect operational anomalies, ensuring that the data used for subsequent anomaly detection model training is generated when the wind turbine is running in a healthy state, and that the data used for subsequent analysis is undamaged data collected under normal operating conditions of the wind turbine, thereby improving the accuracy of the anomaly detection results of this application.

[0053] Furthermore, non-steady-state data is removed from historical operating data. Specifically, during the actual power generation process of wind turbines, due to weather factors such as turbine start-up and shutdown, obstruction, and rapid changes in wind speed, wind turbines may experience alternating non-steady-state and steady-state operating conditions. This results in a large amount of non-steady-state data mixed in with historical operating data. Under non-steady-state conditions, the consistency between the system's input and output parameters cannot be guaranteed. Therefore, this application needs to remove non-steady-state data before establishing an anomaly detection model for operating conditions.

[0054] In one embodiment of this application, to more clearly illustrate the specific implementation process of removing non-steady-state data, the following example illustrates a method for steady-state screening of historical operating data of wind turbines proposed in this application. Figure 2 As shown, the method includes the following steps:

[0055] Step S201: Express the output power of the target wind turbine as an expression that includes the rate of change of power.

[0056] It should be noted that, since the output power of wind turbines exhibits relatively small fluctuations during steady-state processes but large fluctuations during unsteady-state processes, this application selects output power as the characteristic index for steady-state discrimination. Specifically, in the discrimination process, the output power of the wind turbine is first expressed in the form of a formula, which is the expression containing the rate of power change shown in the following formula (1):

[0057]

[0058] Where, p t Let t be the measured power value at time t, μ be the initial true power value, m be the rate of change of power, and ε be the random error of power, with ε following a normal distribution.

[0059] Step S202: Calculate the difference in output power between two adjacent time points.

[0060] Specifically, the difference between steady-state and unsteady-state operation of a wind turbine lies in whether the load response rate is equal to 0, and the load response rate can be reflected by the rate of change of power m. In this embodiment, in order to facilitate the estimation of the value of m, the difference Δp of power between two adjacent moments is calculated, that is, the difference is obtained by subtracting the above formula (1) from the difference between two adjacent moments, as shown in the following formula (2):

[0061] Δp=p t -p t-1 =m+(ε t -ε t-1 (2)

[0062] Step S203: Estimate the rate of change of power using the mean of the sample statistics within the time window.

[0063] Specifically, due to ε t ~N(0,σ 2 The expected value of the statistic Δp is m, i.e., Δp ~ N(m, 2σ). 2 Based on the properties of time series, m can be estimated using the mean of the sample statistics within the time window, which can be achieved using the formula (3) shown below:

[0064]

[0065] Where h is the number of samples within the sliding time window.

[0066] Step S204: Determine the confidence interval in which the true value of the rate of change of power lies through interval estimation.

[0067] Step S205: Determine whether the target wind turbine is in a non-steady-state condition within the time window based on whether the confidence interval does not include zero. If so, remove the non-steady-state data.

[0068] Specifically, to ensure the reliability of the estimation, this embodiment uses interval estimation, as shown in the following formula (4):

[0069]

[0070] in, and These are two statistics at a given significance level α. There exists a confidence level of (1-α) for considering the true value of m to lie within the confidence interval. Therefore, based on whether the confidence interval does not include zero, it can be determined whether the target wind turbine is in a non-steady-state condition within the time window t-1 to t. That is, if the confidence interval does not include zero... If 0 is excluded, the wind turbine is determined to be in a non-steady-state condition during that time period. The operating data during that time period is non-steady-state data, and the non-steady-state data during that time period is removed.

[0071] Therefore, the method for eliminating non-steady-state data in the embodiments of this application eliminates non-steady-state data generated under non-steady-state operating conditions that do not meet strong consistency, reducing the complexity of subsequent modeling and training processes based on historical operating data, reducing the time and effort spent on accurate modeling, and improving the convenience of the anomaly detection method of this application.

[0072] Step S102: Select operating condition characteristic parameters, divide historical operating data into operating conditions through equal interval classification, and cluster the data samples under each operating condition according to sample similarity using Gaussian Mixture Model (GMM). Select benchmark samples from all clusters under each operating condition according to preset screening criteria.

[0073] Among them, operating condition characteristic parameters are operating parameters that directly or indirectly affect the state characteristic parameters of wind turbine units. These parameters can be used to distinguish different operating conditions and can be approximated as boundary conditions for operating condition classification. A Gaussian Mixture Model (GMM) is a linear combination of multiple Gaussian distribution functions, which can accurately quantify phenomena using Gaussian probability density functions.

[0074] In one embodiment of this application, when dividing operating conditions, a set of operating condition characteristics is first established by selecting a corresponding number of operating condition characteristic parameters. As an example, this application selects six operating condition characteristic parameters—wind speed, wind direction, torque, speed, ambient temperature, and load—by comparing the influence of each parameter on the state characteristics of the unit in historical data.

[0075] Then, the filtered historical operating data is traversed, and the maximum and minimum values ​​of each operating condition characteristic parameter within its own range of variation are determined by comparison. The preset operating condition division interval for each characteristic parameter is then obtained. The operating condition division interval for the characteristic parameter can be set according to the characteristics of the current wind turbine, the actual variation range of each parameter, the data distribution characteristics, and the required accuracy of anomaly detection. For example, when higher accuracy is required and the operating condition division is more detailed, the operating condition division interval for each characteristic parameter can be set to be shorter.

[0076] After determining the interval for each characteristic parameter, the equal interval working conditions are divided using the following formula:

[0077] C i =(S i ,S i +ΔS)∩(D i D i +ΔD)∩(M i M i +ΔM)∩(N i N i +ΔN)∩(T i ,T i +ΔT)∩(P i ,P i +ΔP)

[0078]

[0079] Where S is wind speed, D is wind direction, M is torque, N is rotational speed, T is ambient temperature, and P is load. From this formula and the description of the characteristic parameters above, it can be seen that T... min and T max P represents the minimum and maximum ambient temperatures, respectively. min With P maxThese represent the minimum and maximum load values, respectively. ΔT and ΔP represent the ambient temperature and load condition intervals, respectively. Other characteristic parameters can be deduced similarly, and will not be elaborated here.

[0080] As can be seen from the above working condition division formula (5), the i-th working condition Ci divided in this application is determined by the intersection of the division conditions of each operating working condition characteristic parameter. The corresponding working condition is divided according to the intersection of each operating working condition characteristic parameter in its respective division interval.

[0081] It should be noted that, since the division of operating conditions based on equal intervals often results in some invalid operating conditions with no samples or a small number of samples, this embodiment also removes some invalid operating conditions after performing the equal interval division. Specifically, the number of samples in each operating condition after division is detected. If the number of samples is lower than a preset threshold, the operating condition is determined to be invalid, and invalid operating conditions are removed from all the divided operating conditions until the number of remaining valid operating conditions is greater than a preset proportion of the total number of theoretical operating conditions. For example, when the number of remaining valid operating conditions is greater than 30% of the total number of theoretical operating conditions, it is determined that the collected data samples are meaningful relative to the current operating condition division result, thereby completing the operating condition division.

[0082] Furthermore, a benchmark sample is selected from the historical operating data after dividing the operating conditions.

[0083] The baseline sample consists of data samples from wind turbines operating under normal or healthy conditions. However, due to external conditions, the boundaries of normal operating conditions vary under different operating scenarios, making it difficult to uniformly define the benchmark for normal operating conditions. Therefore, this application divides operating conditions and utilizes a Gaussian Mixture Model (GMM) to accurately quantify the characteristics of data. GMM clusters data samples under a single operating condition based on similarity, selecting data from one or more clusters as baseline samples according to predefined screening criteria. GMM combines the advantages of parametric and nonparametric estimation methods, is not limited to a specific probability density function form, and can approximate any continuous distribution with arbitrary precision when the number of sub-models is sufficient. Therefore, this application uses GMM to determine multiple different distributions contained in the historical operating data set, i.e., generating different clusters.

[0084] In a specific implementation, in one embodiment of this application, to more clearly illustrate the specific implementation process of filtering benchmark samples using GMM, the following example illustrates a method for filtering benchmark samples of wind turbine generators proposed in this application. For each single operating condition, benchmark samples can be filtered using this method. For example... Figure 3 As shown, the method includes the following steps:

[0085] Step S301: Estimate the parameters of the Gaussian Mixture Model (GMM) using the expectation-maximization algorithm.

[0086] Specifically, the Expectation-Maximization (EM) algorithm is an iterative algorithm that can be used to estimate the parameters of a probabilistic model containing latent variables. The parameter values ​​are obtained by iteratively finding the maximum likelihood. The objective function of the EM algorithm used in this application is shown in the following formula (6):

[0087]

[0088] In this embodiment, when estimating GMM parameters using the EM algorithm, a rough value of the GMM parameter is first estimated, and then the obtained rough value is used to maximize the likelihood function.

[0089] For example, in this embodiment, the likelihood function of the GMM is first calculated. Assuming that there are N parameters in the GMM model that need to be estimated, the maximum likelihood function of these N parameters is first solved. Then, the GMM parameters are estimated by the EM algorithm, that is, the maximum likelihood function of these N parameters is maximized. Specifically, the initial values ​​of these N parameters can be specified first, and the EM algorithm is iterated until these N parameters or the log-likelihood function converges to obtain the estimated parameter values.

[0090] Step S302: Determine the number of sub-models of the Gaussian mixture model using the Akaike Information Criterion (AIC) to cluster the state categories of the target wind turbine.

[0091] The physical meaning of the number of GMM sub-models is the number of wind turbine state categories. When the number of GMM sub-models is 1, it is equivalent to a Gaussian distribution. Considering the changes in the operating level of the wind turbine in this application, the failures that occur during operation, and the slow aging of the unit under normal conditions, the historical data samples may contain multiple operating modes. Therefore, this application represents the probability distribution as a superposition combination of multiple Gaussian distributions.

[0092] Specifically, in order to determine the number of GMM sub-models, the AIC evaluation criterion is selected in this embodiment. The Akaike Information Criterion (AIC) is an evaluation method based on the concept of entropy, used to balance model complexity and the quality of the fitted data. Its definition is shown in the following formula (7):

[0093] AIC = 2K - 2ln(L)

[0094] Where K is the number of sub-models, and ln(L) represents the log-likelihood function of the model.

[0095] It should be noted that as the number of sub-models increases, the model complexity also increases, and the AIC value will decrease and then increase. Therefore, this embodiment selects the number of sub-models corresponding to the minimum AIC as the optimal number of sub-models for the GMM. Thus, this embodiment determines the number of sub-models by estimating the parameters of the GMM and combining them with the AIC criterion, and clusters the data samples under a single working condition based on similarity. Furthermore, the GMM can then be used to sequentially cluster the data samples under each divided working condition.

[0096] Step S303: Select benchmark samples from all clusters under each working condition according to preset screening criteria.

[0097] In this embodiment, the benchmark is defined as the parameter value corresponding to the highest achievable unit operating efficiency under the current operating boundary conditions. Therefore, the preset screening criterion in this application can be to select the data with the highest operating efficiency as the benchmark sample. Specifically, when screening benchmark samples, for each operating condition, data samples under different clusters determined by GMM can be compared, and the set of data with the highest average operating efficiency can be selected as the benchmark sample. Furthermore, its probability density distribution can be estimated to determine the benchmark interval. Then, the benchmark samples for each divided operating condition are screened sequentially in the same way.

[0098] It should be noted that this application uses historical data reflecting the normal operating status of the unit as the baseline sample. Within the period covered by the historical data, the unit may deviate from its optimal operating state and exhibit some minor anomalies. These minor anomalies are difficult to detect, or even if an anomaly is suspected, it is believed that as long as the anomaly does not continue to develop, this suspected data can still be used as the baseline sample. Therefore, after selecting the baseline sample, this application, based on the trained GMM, treats sample points with a likelihood much lower than the average likelihood as anomalous samples and removes them, thereby further eliminating anomalous data and improving the accuracy of the subsequently constructed anomaly detection model.

[0099] Step S103: Analyze the operating status of the wind turbine under varying operating conditions, construct a benchmark model of the target wind turbine under varying operating conditions based on the conditional variational autoencoder, and train the conditional variational autoencoder using benchmark samples to obtain the trained benchmark model.

[0100] Among them, variable operating conditions refer to the operating scenarios in which the wind turbine is in operation, or deviates from the design operating conditions. This application first analyzes the operating status of the wind turbine under variable operating conditions to determine the method for building a benchmark model under variable operating conditions.

[0101] Specifically, since most of the operational status characteristics of wind turbines are related to boundary conditions and are coupled with each other, changes in the monitorable operating parameters of wind turbines are not only related to the health of the turbine itself but also affected by external operating conditions. Therefore, in one embodiment of this application, when analyzing the variable operating status of wind turbines, historical operating data is first analyzed to construct a... Figure 4 The diagram shows the relationship between characteristic indices and boundary variables of a wind turbine under varying operating conditions. Figure 4 In this diagram, X represents directly monitorable primary parameters, which can be obtained directly from sensor points located at various positions on the wind turbine, including temperature, wind speed, and vibration. Y represents secondary performance indicators, which need to be obtained through soft measurement of parameters. Secondary performance indicators are more sensitive to changes in the unit's condition and can quickly locate the location of anomalies. C represents external operating condition variables, including environmental parameters and the unit's rotational speed. Z represents the unit's state variables, reflecting the unit's own health condition and original design capabilities.

[0102] based on Figure 4 The relationship shown represents the operating state model of the wind turbine under varying operating conditions as a conditional probability distribution, expressed as logp. θ (x,y|c) indicates the predicted values ​​of characteristic indicators X and Y under given operating condition variable C. After determining the characteristic indicators to be monitored, the parameters of the probability model are identified using the screened data samples, and a representative maximum likelihood estimate is selected for solution, as shown in the following formula (8):

[0103]

[0104] Where N represents the number of training samples, and for the sake of brevity and without ambiguity, x is referred to as such in this application. (i) Let x be the variable. Since operating parameters and performance indicators are also affected by unit state variables Z in addition to operating condition variables, the probabilistic model contains both observed and latent variables. Directly solving for logp... θ (x,y|c) is relatively difficult to learn directly from data and is not suitable for learning conditional probability distributions directly from data. Therefore, this application, based on Bayes' theorem, uses the conditional probability distribution of latent variables to realize the probability distribution of observable variables, that is, logp θ (x,y|c) can be expressed in the form shown in the following formula.

[0105] logp θ (x,y|c)=logp θ (x,y|z,c)+logp θ (z|c)-logp θ (z|x,y,c) (9)

[0106] As can be seen from this formula, when the characteristic indicators of the unit are coupled with each other, there are complex nonlinear relationships between the variables, and the posterior probability of the latent variable z is logp. θ (z|x,y,c) is not a simple probability distribution. Especially when the conditional probability logp in the equation... θ When the parameters in (z|x,y,c) are obtained through neural network calculation, the EM algorithm is no longer applicable. Therefore, this application uses an improved conditionally variable autoencoder to establish a baseline model of the unit under varying operating conditions.

[0107] Therefore, this application, through analysis of the operating state of wind turbine units under varying operating conditions, determines that a baseline model based on a conditional variational autoencoder is more suitable for use under varying operating conditions. To more clearly explain the principle behind this application's use of an improved conditional variational autoencoder to establish a baseline model of the unit under varying operating conditions, the variational autoencoder is first introduced below:

[0108] A Variational Autoencoder (VAE) is a deep learning generative model that integrates the advantages of deep learning on the basis of probabilistic models. It can autonomously learn the probability distribution that data follows and generate similar data. The network structure of a VAE consists of two parts: an encoder and a decoder.

[0109] The encoder, also called the inference network, functions to transmit data through p... θ (z|x) maps the original data x to a low-dimensional latent space, encoding it into latent variables z, thus achieving feature extraction and dimensionality reduction. The decoder, also called a generative network, is responsible for extracting features from the latent variable distribution p. θ (z) Sample z, through p θ (x|z) Reconstructs the original data x.

[0110] decoder p θ Integrating multi-layer neural networks (x|z) can improve the ability to express complex nonlinear relationships, but the posterior probability p of the latent variable z... θ (z|x) would thus become more complex, making it impossible to obtain an analytical form of the distribution. Therefore, VAE utilizes variational inference to construct an optimizable q. φ (z|x) is used to approximate the true posterior probability p in the encoder. θ (z|x). The log-likelihood of the observed variable x can be written as:

[0111] logp θ (x)=D KL (q φ (z|x)||p θ (z|x))+L ELBO (θ,φ;x) (10)

[0112] Among them, D KL Let represent the KL divergence, used to measure the similarity between two probability distributions in the same event space; its value is always non-negative. Therefore, the problem of maximizing the log-likelihood of the observed variable x is equivalent to maximizing the variational lower bound L. ELBO (θ,φ;x), as shown in equation (11).

[0113] logp θ (x)≥L ELBO (θ,φ;x) (11)

[0114] The equality of formula (11) holds if the variational approximate posterior probability equals the true posterior probability. ELBO (θ,φ;x) can be further expanded as shown in formula (12).

[0115]

[0116] Formula (12) contains two terms: the first term is the reconstruction error of the observed variable x, which is approximated by the posterior distribution q. φ (z|x) Sample the latent variable z and calculate the log-likelihood logp θ (x|z) is used to describe the difference between the generated sample and the real sample; the second term is the approximate posterior distribution q of the latent variable. φ (z|x) and the prior distribution p θ The KL dispersion among (z) constrains the posterior distribution to approximate the prior distribution and can be considered as the regularization coefficient of the latent variable z. The VAE uses the SGVB training algorithm to optimize the parameters Φ and θ, satisfying the following two conditions: first, making the training data and reconstructed data as equal as possible; second, ensuring that the posterior distribution q of z... φ (z|x) approximates the prior distribution p θ (z).

[0117] VAEs typically assume that the posterior distribution of z is q. φ (z|x) follows a Gaussian distribution, with a prior distribution p. θ (z) follows a standard normal distribution, and the likelihood p θ (x|z) Choose a multivariate Gaussian distribution or a Bernoulli distribution based on the characteristics of the variables. Directly using Monte Carlo sampling to estimate the parameters of the model will produce a large variance. In order to reduce the variance, the latent variables are transformed by reparameter transformation, which is represented by a differentiable equation and a random variable, as shown in formula (12).

[0118] z=g(x,y,c,ε)withε~p(ε) (13)

[0119] When the prior probability satisfies p θ (z)=N(0,I), qφ (z|x)=N(z;μ,σ 2 When condition I) is met, the expression for the variational lower bound is equivalent to the following formula (14):

[0120]

[0121] In this embodiment of the application, since the operating state model of the wind turbine is based on the conditional probability distribution logp of characteristic indices X and Y under the condition variable C, θ (x,y|c) is used for monitoring. Therefore, based on the structure of the standard VAE mentioned above, this application introduces conditional variables to improve the original network structure and establishes a Conditional Variational Auto-Encoder (CVAE) model to predict baseline data under varying operating conditions.

[0122] Specifically, when constructing a baseline model based on a conditional variational autoencoder, as one possible implementation, this application embodiment constructs the following... Figure 5 The network structure of CVAE is shown below. Figure 5 As shown, the model includes an encoding module 10, a decoding module 20, and a module 30 for acquiring and transmitting operating conditions (C). The objective of CVAE is to solve for the conditional log-likelihood logp. θ The value of the parameter is maximized when (x,y|c). Under varying operating conditions, first convert the equation logp... θ (x) expands to the conditional log-likelihood logp θ (x, y|c) can be written as the following formula:

[0123] logp θ (x,y|c)=D KL (q φ (z|x,y,c)||p θ (z|x,y,c))+L ELBO (θ,φ;x,y,c) (15)

[0124] Then, based on the fundamental principles of VAE, the variational lower bound L of the conditional log-likelihood is... ELBO (θ,φ;x,y,c) expands to the following expression:

[0125]

[0126] In this embodiment, it is assumed that the prior distribution of the latent variables follows an isotropic multivariate Gaussian distribution, and the prior probability does not involve parameters, as shown in the following formula:

[0127] p(z)=N(0,I) (17)

[0128] Conditional likelihood p of the observed variable θ (x,y|z,c) follows a multivariate Gaussian distribution, as shown in the following formula (18):

[0129] p θ (x,y|z,c)=N(μ θ (z,c),σ θ (z,c) 2 I) (18)

[0130] The parameters are obtained from the output of the multi-layer neural network in the encoder, σ θ (z,c) is assumed to be a constant, μ θ (z,c)=W θ [z,c]+b θ , σ θ (z,c) = const.

[0131] Because the true posterior probability p θ The form (z|x,y,c) is a multivariate Gaussian model with diagonal variance. To simplify the estimation of the posterior probability, a multivariate Gaussian distribution with diagonal variance is also used to represent the encoder distribution q. φ (z|x,y,c), as shown in the following formula.

[0132] q φ (z|x,y,c)=N(μ φ (x,y,c),σ φ (x,y,c) 2 I) (19)

[0133] The variational approximation distribution after reparameterization is shown in the following formula:

[0134] z = μ φ (x,y,c)+σ φ (x,y,c)⊙εwithε~N(0,I) (20)

[0135] The parameters can also be calculated by a neural network, as shown in the following formula:

[0136]

[0137] Therefore, this application constructs a benchmark model of the target wind turbine under varying operating conditions based on a conditional variational autoencoder. Then, the CVAE model is trained using the benchmark samples obtained in step S102 for each operating condition. After training, the benchmark model of the wind turbine under varying operating conditions can be obtained.

[0138] Step 104: Calculate the reconstruction probability as a feature index for anomaly detection using the trained benchmark model, and obtain the real-time operation data of the target wind turbine. Based on the reconstruction probability and the deviation between the actual value and the reconstruction value of the real-time operation data, locate the abnormal operation data.

[0139] The reconstruction probability is the reconstruction error term in the loss function of the CVAE model. It should be noted that the CVAE model reconstructs the probability distribution of variables, not the variables themselves. The reconstructed data contains more information about the variables, including the posterior distribution of latent variables and the generated likelihood. For the CVAE benchmark model, the lower the reconstruction probability of the test data, the greater the difference between the data and the benchmark. Therefore, this application selects the reconstruction probability as a feature indicator for anomaly detection.

[0140] Specifically, this application first calculates the reconstruction probability as a feature index for anomaly detection using a trained benchmark model. As one possible implementation, when calculating the reconstruction probability, test samples are first acquired. The encoder of a conditional variational autoencoder obtains the first and second parameters from the Gaussian distribution of each latent variable, and a preset number of sample points are sampled for each latent variable. Then, the decoder of the conditional variational autoencoder calculates the third and fourth parameters from the likelihood distribution of the reconstruction variable corresponding to each latent variable. Finally, based on the third and fourth parameters, the average log-likelihood of the test samples under the latent variable conditions is calculated.

[0141] Specifically, the test samples can be sample data selected from the historical operating data of the target wind turbine to calculate the reconfiguration probability. After inputting the test sample X into the CVAE baseline model, the encoder of the CVAE obtains the first parameter μ in the Gaussian distribution of the latent variables. φ Second parameter σ φ The latent variable Z is sampled at L points. Then, the decoder calculates the third parameter μ of the likelihood distribution of the reconstructed variable corresponding to each latent variable Z. θ and the fourth parameter σ θ Finally, the reconstruction probability is obtained by averaging the log-likelihoods of the test samples under the latent variable condition. The specific calculation method can be expressed by the following formula:

[0142]

[0143] Furthermore, after generating a baseline model under varying operating conditions and calculating the reconfiguration probability, real-time operating data of the target wind turbine to be detected is obtained, and the difference between the current data and the baseline data is compared to identify anomalies.

[0144] One possible implementation involves identifying anomalous running data based on the reconstruction probability and the deviation between the actual and reconstructed values ​​of the real-time running data. This includes the following steps: First, the reconstruction probability distribution of the training samples is statistically estimated using KDE (Kernel Density Estimation). The lower bound of the confidence interval corresponding to the confidence level equaling a preset value is used as the threshold for anomaly detection. Then, based on the anomaly detection threshold, it is determined whether the reconstruction probability of the real-time running data is anomalous. If anomaly in the reconstruction probability is detected, the deviation of each parameter in the real-time running data is calculated using the following formula:

[0145]

[0146] Among them, v k x represents the standardized coefficient. k This represents the actual value of parameter k. This represents the reconstructed value of parameter k. Finally, the deviation of each parameter is compared to determine the parameter that caused the anomaly.

[0147] Specifically, kernel density estimation (KDE) is used to statistically analyze the reconstruction probability distribution of the training samples. The lower bound of the confidence interval corresponding to a confidence level of 95% is used as the threshold for anomaly detection. The specific calculation formula is as follows:

[0148]

[0149] The training samples can be sample data extracted from the acquired real-time operating data to calculate the threshold for anomaly detection. In wind turbine units, the mutual coupling between parameters such as wind speed and rotational speed means that disturbances in local parameters can be transmitted to other parameters within the system. Furthermore, under the control and regulation of the system, small deviations of some parameters that were abnormal in the early stages can be pulled back to normal levels, affecting the location of parameter anomalies.

[0150] Therefore, this application employs anomaly detection based on reconfiguration probability to detect abnormal changes in local system units of the unit in advance. Furthermore, it calculates the difference vector between the reconfigured output and the original input parameters, and uses the relative deviation between the two to further locate parameter anomalies. The formula for calculating the parameter deviation is as follows:

[0151]

[0152] Among them, v k denoted by , x represents the standardization coefficient, which is obtained statistically from the absolute deviation between the reconstructed output and the input signal from the training samples. k and These represent the actual value and the reconstructed value of parameter k, respectively.

[0153] In this embodiment, after obtaining the anomaly detection threshold by estimating the reconstruction probability distribution of the KDE statistical training samples through kernel density estimation, the reconstruction probability of the real-time operating data is calculated, and the reconstruction probability of the real-time operating data is compared with the anomaly detection threshold. Once an anomaly in the reconstruction probability of the real-time operating data is detected, it is determined that the target wind turbine is currently in an abnormal state. Further, the deviation of each operating parameter in the operating data, such as the wind turbine torque, is calculated using the above formula (24), and the parameter anomaly is further located by comparing the deviation of each parameter in the system. For example, the deviation of each operating parameter is compared with a preset deviation threshold. When the deviation of a certain operating parameter is greater than the preset deviation threshold, it is determined that the operating parameter is abnormal.

[0154] Therefore, this application reconstructs the probability. As a feature index for anomaly detection, reconstruction probability not only considers the difference between the reconstructed and original inputs but also the impact of latent variable variance on the reconstruction results, thus possessing the potential to detect abnormal data fluctuations. Furthermore, reconstruction probability measures the deviation of the current overall state from the baseline, making it more sensitive to anomaly detection in multi-parameter coupled systems. Therefore, this application constructs a baseline model under varying operating conditions based on a conditional variational autoencoder. By comparing real-time operating data with the baseline data of the current scenario determined by the baseline model using reconstruction probability, accurate anomaly detection results under varying operating conditions can be obtained. This not only allows for timely and accurate detection of whether the wind turbine's operating state is abnormal but also, by calculating the deviation between the actual and reconstructed values, precisely identifies the specific parameters that caused the anomaly.

[0155] In summary, the wind turbine operation data anomaly detection method of this application first initiates a data cleaning process, establishing steady-state discrimination, operating condition division, and benchmark screening models. Invalid anomaly data, non-steady-state data, and a small number of minor anomaly data points mixed in with the historical data of the wind turbine are removed. The cleaned data is used for training the subsequent anomaly detection model, which can improve the accuracy of the trained anomaly detection model in detecting anomaly data. Then, considering the dynamic complexity of wind turbine state information, the correlation between multiple parameters in the data is analyzed in a data-driven manner. A variable autoencoder model is used to reconstruct the data under varying operating conditions, establishing a multi-feature parameter fusion anomaly detection model to achieve real-time dynamic assessment of the health status of the wind turbine under varying operating conditions. Therefore, this method can establish an accurate benchmark model in scenarios where operating conditions change and deviate from the design conditions, which is beneficial for timely and accurate detection of abnormal parameters under varying operating conditions. This improves the accuracy of assessing the operating status of the wind turbine under varying operating conditions and can accurately locate specific abnormal parameters, allowing staff to obtain more detailed anomaly information and facilitate timely troubleshooting.

[0156] To more clearly illustrate the flow of the wind turbine operating data anomaly detection method according to the embodiments of this application, a detailed description is provided below using a specific embodiment of wind turbine operating data anomaly detection. Figure 6 This is a flowchart illustrating a specific method for detecting abnormal operating data of a wind turbine generator according to an embodiment of this application. Figure 6 As shown, the method includes the following steps:

[0157] Step S601: Identify and filter out normal operation data from historical operation data.

[0158] Step S602: Divide the working conditions based on the working condition division method of equal interval classification.

[0159] Step S603: Filter the baseline sample data.

[0160] Steps S601 to S603 constitute the historical data cleaning process. To address the issues of invalid, non-steady-state, and a small number of abnormal data points mixed in with historical data, steady-state discrimination, working condition division, and benchmark screening models were studied to clean the data. The cleaned data is used for training the subsequent anomaly detection model.

[0161] Step S604: Analyze the variable operating status of the wind turbine.

[0162] Step S605: Construct a wind turbine variable operating condition benchmark model based on a conditional autoencoder.

[0163] Step S606: Perform anomaly detection based on reconstruction probability.

[0164] Steps S604 to S606 constitute the variable operating condition anomaly detection process. In view of the dynamic complexity of wind turbine status information, the correlation between multiple parameters in the data is analyzed in a data-driven manner. The variable autoencoder model is used to reconstruct the data of variable operating conditions, and a multi-feature parameter fusion anomaly detection model is established to realize the real-time dynamic assessment of the health status of wind turbines under variable operating conditions.

[0165] It should be noted that, in this embodiment, the specific implementation of each step can be referred to the description of the above embodiment, and the specific implementation is similar, so it will not be repeated here.

[0166] To achieve the above embodiments, this application also proposes a wind turbine operating data anomaly detection system. Figure 7 This is a schematic diagram of the structure of a wind turbine operation data anomaly detection system proposed in an embodiment of this application, as shown below. Figure 7 As shown, the system includes an acquisition module 100, a filtering module 200, a construction module 300, and a detection module 400.

[0167] The acquisition module 100 is used to acquire the historical operating data of the target wind turbine to be detected, and to remove abnormal data and non-steady-state data that do not conform to the output characteristics of the wind turbine from the historical operating data.

[0168] The filtering module 200 is used to select operating condition characteristic parameters, divide historical operating data into operating conditions through equal interval classification, and cluster the data samples under each operating condition according to sample similarity using Gaussian Mixture Model (GMM). Based on preset filtering criteria, benchmark samples are selected from all clusters under each operating condition.

[0169] Module 300 is used to analyze the operating status of wind turbines under varying operating conditions. It constructs a benchmark model of the target wind turbine under varying operating conditions based on a conditional variational autoencoder, and trains the conditional variational autoencoder using benchmark samples to obtain the trained benchmark model.

[0170] The detection module 400 is used to calculate the reconstruction probability as a feature index for anomaly detection through the trained benchmark model, and to acquire the real-time operation data of the target wind turbine. Based on the reconstruction probability and the deviation between the actual value and the reconstruction value of the real-time operation data, the abnormal operation data is located.

[0171] Optionally, in one embodiment of this application, the acquisition module 100 is specifically used to: express the output power of the target wind turbine as an expression that includes the rate of change of power; calculate the difference in output power between two adjacent moments; estimate the rate of change of power by the mean of the sample statistics within the time window; determine the confidence interval in which the true value of the rate of change of power lies by interval estimation; and determine whether the target wind turbine is in a non-steady-state condition within the time window based on whether the confidence interval does not include zero. If so, the non-steady-state data is removed.

[0172] Optionally, in one embodiment of this application, the selected operating condition characteristic parameters include: wind speed, wind direction, torque, rotational speed and ambient temperature. The filtering module 200 is specifically used to: determine the maximum and minimum values ​​of each operating condition characteristic parameter within its own range of variation, and obtain the preset operating condition division interval for each operating condition characteristic parameter.

[0173] The following formula is used to divide the operating conditions into equal intervals:

[0174] C i =(S i ,S i +ΔS)∩(D i D i +ΔD)∩(M i M i +ΔM)∩(N i N i+ΔN)∩(T i ,T i +ΔT)∩(P i ,P i +ΔP)

[0175] stS i <<S min ,S i +ΔS>>S max D i <<D min D i +ΔD>>D max M i <<M min M i +ΔM>>M max N i <<N min N i +ΔN>>N max ,T i <<T min ,T i +ΔT>>T max ,P i <<P min ,P i +ΔP>>P max

[0176] Where S is wind speed, D is wind direction, M is torque, N is rotational speed, T is ambient temperature, and P is load; invalid working conditions are removed from all the divided working conditions until the number of remaining valid working conditions is greater than the preset proportion of the total number of theoretical working conditions.

[0177] Optionally, in one embodiment of this application, the screening module 200 is further configured to: estimate the parameters of the Gaussian Mixture Model (GMM) using the Expectation-Maximization (EMA) algorithm; determine the number of sub-models of the GMM using the Akaike Information Criterion (AIC) to cluster the state categories of the target wind turbine; compare the data samples under each cluster and select the target data group with the highest average operating efficiency as the benchmark sample; and remove sample points with a likelihood much smaller than the average likelihood using the GMM.

[0178] Optionally, in one embodiment of this application, the detection module 400 is specifically used to: acquire test samples; acquire the first and second parameters in the Gaussian distribution of each latent variable through the encoder of the conditional variational autoencoder; and sample a preset number of sample points for each latent variable; calculate the third and fourth parameters in the likelihood distribution of the reconstructed variable corresponding to each latent variable through the decoder of the conditional variational autoencoder; and, based on the third and fourth parameters, calculate the average value of the log-likelihood of the test samples under the latent variable conditions.

[0179] Optionally, in one embodiment of this application, the detection module 400 is further configured to: estimate the reconstruction probability distribution of the KDE statistical training samples by kernel density estimation, and use the lower limit of the confidence interval corresponding to the confidence level being equal to a preset value as the threshold for anomaly detection; determine whether the reconstruction probability of the real-time running data is abnormal based on the anomaly detection threshold, and calculate the deviation of each parameter in the real-time running data using the following formula if an abnormal reconstruction probability is detected:

[0180]

[0181] Among them, v k x represents the standardized coefficient. k This represents the actual value of parameter k. This represents the reconstructed value of parameter k; by comparing the deviation of each parameter, the parameter that caused the anomaly is determined.

[0182] It should be noted that the explanation of the aforementioned embodiment of the method for detecting abnormal operating data of wind turbine units also applies to the system of this embodiment, and will not be repeated here.

[0183] In summary, the wind turbine operation data anomaly detection system of this application first initiates a data cleaning process. The cleaned data is used to train the subsequent anomaly detection model, which can improve the accuracy of the trained anomaly detection model in detecting abnormal data. Then, considering the dynamic complexity of wind turbine state information, the system analyzes the correlation between multiple parameters in the data in a data-driven manner, and uses a variable autoencoder model to reconstruct the data under varying operating conditions, establishing a multi-feature parameter fusion anomaly detection model to achieve real-time dynamic assessment of the health status of the wind turbine under varying operating conditions. Thus, the system can establish an accurate benchmark model in scenarios where operating conditions change and deviate from the design conditions, which is beneficial for timely and accurate detection of abnormal parameters under varying operating conditions, improving the accuracy of wind turbine operation status assessment under varying operating conditions, and accurately locating specific abnormal parameters.

[0184] To implement the above embodiments, this application also proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the wind turbine operating data anomaly detection method as described in any of the above embodiments.

[0185] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. 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 different embodiments or examples.

[0186] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0187] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0188] 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). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0189] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0190] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0191] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0192] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A method for detecting abnormal operating data of a wind turbine generator, characterized in that, Includes the following steps: Obtain historical operating data of the target wind turbine to be tested, and remove abnormal data and non-steady-state data that do not conform to the output characteristics of the wind turbine from the historical operating data; Select operating condition characteristic parameters, divide the historical operating data into operating conditions through equal interval classification, and cluster the data samples under each operating condition according to sample similarity using Gaussian Mixture Model (GMM). Select benchmark samples from all clusters under each operating condition according to preset screening criteria. The operating status of wind turbines under varying operating conditions is analyzed. A baseline model of the target wind turbine under varying operating conditions is constructed based on a conditional variational autoencoder. The conditional variational autoencoder is then trained using the baseline samples to obtain the trained baseline model. The reconstruction probability is calculated using the trained benchmark model as a feature index for anomaly detection. Real-time operating data of the target wind turbine is obtained, and abnormal operating data is located based on the reconstruction probability and the deviation between the actual value and the reconstruction value of the real-time operating data. The method of clustering data samples under each working condition based on sample similarity using a Gaussian Mixture Model (GMM) includes: The parameters of the Gaussian mixture model (GMM) are estimated using the expectation-maximization algorithm. The number of sub-models of the Gaussian mixture model is determined by the Akaike Information Criterion (AIC) in order to cluster the state categories of the target wind turbine. The step of selecting benchmark samples from all clusters under each working condition according to preset screening criteria includes: Compare the data samples under each cluster and select the target data group with the highest average running efficiency as the benchmark sample; The Gaussian Mixture Model (GMM) is used to remove sample points whose likelihood is much smaller than the average likelihood.

2. The method according to claim 1, characterized in that, The process of removing non-steady-state data from the historical operating data includes: The output power of the target wind turbine is expressed as an expression that includes the rate of change of power; Calculate the difference in output power between two adjacent time points; The rate of change of the power is estimated by the mean of the sample statistics within the time window; The confidence interval in which the true value of the rate of change of the power lies is determined by interval estimation; Based on whether the confidence interval does not include zero, it is determined whether the target wind turbine is in a non-steady-state condition within the time window. If so, the non-steady-state data is removed.

3. The method according to claim 1 or 2, characterized in that, The operating condition characteristic parameters include: wind speed, wind direction, torque, speed, ambient temperature, and load. The process of classifying the historical operating data into operating conditions through equal-interval classification includes: Determine the maximum and minimum values ​​of each operating condition characteristic parameter within its own range of variation, and obtain the preset operating condition division interval for each operating condition characteristic parameter; The following formula is used to divide the operating conditions into equal intervals: in, It's wind speed. It's the wind direction. It's torque. It's the rotational speed. It is the ambient temperature. It is a load; Eliminate invalid working conditions from all the divided working conditions until the number of remaining valid working conditions is greater than the preset proportion of the total number of theoretical working conditions.

4. The method according to claim 1, characterized in that, The calculation of the reconstruction probability using the trained baseline model includes: Obtain test samples, and use the encoder of the conditional variational autoencoder to obtain the first and second parameters in the Gaussian distribution of each latent variable, and sample a preset number of sample points for each latent variable; The third and fourth parameters in the likelihood distribution of the reconstructed variable corresponding to each latent variable are calculated by the decoder of the conditional variational autoencoder. Based on the third and fourth parameters, the average log-likelihood of the test sample under the latent variable condition is calculated.

5. The method according to claim 1, characterized in that, Based on the reconstruction probability and the deviation between the actual value and the reconstructed value of the real-time running data, abnormal running data is located, including: The reconstruction probability distribution of KDE statistical training samples is estimated by kernel density, and the lower limit of the confidence interval corresponding to the confidence level equal to the preset value is used as the threshold for anomaly detection. Based on the anomaly detection threshold, it is determined whether the reconstruction probability of the real-time running data is abnormal. If an abnormal reconstruction probability is detected, the deviation of each parameter in the real-time running data is calculated using the following formula: in, Represents the standardized coefficient. Indicates parameters k The actual value, Indicates parameters k The reconstructed value; Compare the deviation of each parameter to identify the parameter causing the anomaly.

6. A system for detecting abnormal operating data of a wind turbine generator, characterized in that, include: The acquisition module is used to acquire the historical operating data of the target wind turbine to be detected, and to remove abnormal data and non-steady-state data that do not conform to the output characteristics of the wind turbine from the historical operating data; The filtering module is used to select operating condition characteristic parameters, divide the historical operating data into operating conditions through equal interval classification, and cluster the data samples under each operating condition according to sample similarity using Gaussian Mixture Model (GMM). Based on the preset filtering criteria, the module selects benchmark samples from all clusters under each operating condition. The module is used to analyze the operating status of wind turbines under varying operating conditions, construct a benchmark model of the target wind turbine under varying operating conditions based on a conditional variational autoencoder, and train the conditional variational autoencoder using the benchmark samples to obtain the trained benchmark model. The detection module is used to calculate the reconstruction probability as a feature index for anomaly detection through the trained benchmark model, and to obtain the real-time operation data of the target wind turbine. Based on the reconstruction probability and the deviation between the actual value and the reconstruction value of the real-time operation data, the abnormal operation data is located. The filtering module is used to estimate the parameters of the Gaussian Mixture Model (GMM) using the expectation-maximization algorithm. The number of sub-models of the Gaussian mixture model is determined by the Akaike Information Criterion (AIC) in order to cluster the state categories of the target wind turbine. The step of selecting benchmark samples from all clusters under each working condition according to preset screening criteria includes: Compare the data samples under each cluster and select the target data group with the highest average running efficiency as the benchmark sample; The Gaussian Mixture Model (GMM) is used to remove sample points whose likelihood is much smaller than the average likelihood.

7. The system according to claim 6, characterized in that, The acquisition module is specifically used for: The output power of the target wind turbine is expressed as an expression that includes the rate of change of power; Calculate the difference in output power between two adjacent time points; The rate of change of the power is estimated by the mean of the sample statistics within the time window; The confidence interval in which the true value of the rate of change of the power lies is determined by interval estimation; Based on whether the confidence interval does not include zero, it is determined whether the target wind turbine is in a non-steady-state condition within the time window. If so, the non-steady-state data is removed.

8. The system according to claim 6 or 7, characterized in that, The operating condition characteristic parameters include: wind speed, wind direction, torque, rotational speed, ambient temperature, and load. The screening module is specifically used for: Determine the maximum and minimum values ​​of each operating condition characteristic parameter within its own range of variation, and obtain the preset operating condition division interval for each operating condition characteristic parameter; The following formula is used to divide the operating conditions into equal intervals: in, It's wind speed. It's the wind direction. It's torque. It's the rotational speed. It is the ambient temperature. It is a load; Eliminate invalid working conditions from all the divided working conditions until the number of remaining valid working conditions is greater than the preset proportion of the total number of theoretical working conditions.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the method for detecting abnormal operating data of wind turbine units as described in any one of claims 1-5.

Images