Wind turbine fault diagnosis method, system and electronic device based on SCADA data

Abstract

The present invention provides a wind turbine fault diagnosis method, system and electronic device based on SCADA data. The method includes: obtaining a plurality of sample operation data; training at least part of the sample operation data among the plurality of sample operation data through a support vector machine algorithm to determine support sample operation data, where the support sample operation data is boundary sample operation data for distinguishing positive sample operation data and negative sample operation data; determining generation parameters according to the distances between the support sample operation data and K nearest neighbor sample operation data; generating a plurality of new sample operation data according to the generation parameters to obtain training data for a fan fault diagnosis model; performing feature fusion on the training data to obtain fused training data; and training the fused training data through a support vector machine algorithm to obtain a fan fault diagnosis model. This method solves the problem of complex establishment of detection models in existing detection methods.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power generation, and in particular, to a method, a system and an electronic device for fault diagnosis of a wind turbine based on SCADA data. Background Art

[0002] With the increase in the installed capacity of wind power generation, the normal operation of wind turbines is crucial for the reliability of wind power generation.

[0003] In the prior art, some wind farms have deployed a condition monitoring system for predictive maintenance. This condition monitoring system relies on high-frequency sensing devices to detect faults in wind turbines, resulting in high costs for fault detection. In order to reduce the operation and maintenance costs, a fault detection method based on the operating state of equipment is widely used in the fault detection of wind turbines in wind farms. This detection method mainly realizes the fault detection of wind turbines based on the vibration signals collected from the main drive system of the wind turbines. However, the analysis of vibration signals requires a large amount of experimental data and professional knowledge support, and the establishment of the detection model needs to combine the structure, working principle and the relationship between various components of the wind turbine, resulting in a complex establishment of the detection model. Summary of the Invention

[0004] In view of this, the present invention provides a method, a system and an electronic device for fault diagnosis of a wind turbine based on SCADA data. The detection model of the wind turbine fault trained by this solution can be obtained by training based on historical operation data, solving the problem of complex establishment of the detection model in the existing detection method.

[0005] According to one aspect of the present invention, an embodiment of the present invention provides a method for fault diagnosis of a wind turbine based on SCADA data, the method comprising:

[0006] Obtain multiple sample operation data, where the sample operation data refers to the operation data corresponding to the calibrated fan fault status. The multiple sample operation data includes multiple positive sample operation data and multiple negative sample operation data. The fan fault status corresponding to the positive sample operation data is normal, and the fan fault status corresponding to the negative sample operation data is faulty; Train at least part of the sample operation data in the multiple sample operation data through the support vector machine algorithm to determine the support sample operation data, where the support sample operation data is the boundary sample operation data for distinguishing positive sample operation data and negative sample operation data; Determine the generation parameter according to the distance between the support sample operation data and the K nearest neighbor sample operation data, where K is an integer greater than 1; Generate multiple new sample operation data according to the generation parameter to obtain the training data of the fan fault diagnosis model, where the training data includes the multiple sample operation data and the multiple new sample operation data; Perform feature fusion on the training data to obtain the fused training data; Train the fused training data through the support vector machine algorithm to obtain the fan fault diagnosis model.

[0007] Optionally, when the number of positive sample operation data in the multiple sample operation data is less than the number of negative sample operation data, the determining the generation parameter according to the distance between the support sample operation data and the K nearest neighbor sample operation data includes:

[0008] Determine the positive support sample operation data in the support sample operation data; Calculate the first distance between each positive support sample operation data and each positive sample operation data in the corresponding K nearest neighbor sample operation data; Determine the generation parameter corresponding to each positive support sample operation data according to the first distance.

[0009] Optionally, the determining the generation parameter corresponding to each positive support sample operation data according to the first distance and the second distance includes:

[0010] Calculate the mean value of the first distance to obtain the generation parameter.

[0011] Optionally, the generating multiple new sample operation data according to the generation parameter includes:

[0012] When the number of positive sample operation data in the K nearest neighbor sample operation data corresponding to the positive support sample operation data is less than the number of negative sample operation data, the new sample operation data is:

[0013] X new = x SV+ + δ(x nn+ - x SV+ ) + αγ 0 ;

[0014] When the number of positive sample running data among the K nearest neighbor sample running data corresponding to the positive support sample running data is more than the number of negative sample running data, the new sample running data is:

[0015] X new = x SV+ - δ(x nn+ - x SV+ ) + αγ 0 ;

[0016] where x SV+ is any one of the positive support sample running data, δ is a random number in the interval (0, 1), x nn+ is any one of the positive sample running data among the K nearest neighbor sample running data corresponding to x SV+ , α is a random number in the interval (0, 1), and γ 0 is the generation parameter.

[0017] Optionally, when the number of positive sample running data among the multiple sample running data is more than the number of negative sample running data, determining the generation parameter according to the distance between the support sample running data and the K nearest neighbor sample running data includes: determining the negative support sample running data in the support sample running data; calculating the third distance between each negative support sample running data and each negative sample running data among the corresponding K nearest neighbor sample running data; and determining the generation parameter corresponding to each negative support sample running data according to the third distance.

[0018] Determining the generation parameter corresponding to each negative support sample running data according to the third distance includes:

[0019] calculating the mean value of the third distance to obtain the generation parameter.

[0020] Optionally, generating multiple new sample running data according to the generation parameter includes:

[0021] When the number of negative sample running data among the K nearest neighbor sample running data corresponding to the negative support sample running data is less than the number of positive sample running data, the new sample running data is:

[0022] X new = x SV- + δ(x nn- - x SV- ) + βγ 0 ;

[0023] When the number of negative sample running data among the K nearest neighbor sample running data corresponding to the negative support sample running data is more than the number of positive sample running data, the new sample running data is:

[0024] Xnew = x SV- -δ(x nn- - x SV- ) + βγ 0 ;

[0025] Wherein, x SV- is any one of the negative support sample running data, δ is a random number within the interval (0, 1), x nn- is any one of the negative sample running data among the K nearest neighbor sample running data corresponding to x SV- , β is a random number within the interval (0, 1), and γ 0 is the generation parameter.

[0026] Optionally, the feature fusion of the training data to obtain the fused training data includes:

[0027] Determine the neighborhood set corresponding to each training data, where the neighborhood set includes the N training data closest to the corresponding training data; calculate the linear weight coefficients of each training data and the N training data in the corresponding neighborhood set; optimize the preset objective function based on the linear weight coefficients until the preset objective function converges to obtain the fused training data; the preset objective function includes the sum of regularization and error.

[0028] According to another aspect of the present invention, the present invention provides a wind turbine fault diagnosis system based on SCADA data, and the system includes:

[0029] A data acquisition module for acquiring wind turbine operation data;

[0030] A feature fusion module for performing feature fusion on the wind turbine operation data to obtain fused data;

[0031] A fault diagnosis module for inputting the fused data into a wind turbine fault diagnosis model to diagnose the wind turbine; wherein, the wind turbine fault diagnosis model is obtained according to the above-mentioned wind turbine fault diagnosis method based on SCADA data.

[0032] According to yet another aspect of the present invention, the present invention provides an electronic device, which includes a memory and a processor, and the memory stores a calculation program. When the calculation program is executed by the processor, the processor implements the above-mentioned wind turbine fault diagnosis method based on SCADA data.

[0033] The present invention has the following beneficial effects compared with the prior art:

[0034] (1) The wind turbine fault detection model trained by this method can be obtained by training based on historical operation data, which solves the problem of complex establishment of the detection model in the existing detection methods.

[0035] (2) During the model training process, new sample operation data is generated based on the obtained historical operation data, increasing the number of samples and making the trained fault detection model more reliable.

[0036] (3) When generating new sample operation data, the boundary between positive sample operation data and negative sample operation data is determined by the support vector machine method, and new sample operation data is generated based on the support sample operation data on the boundary, ensuring the reliability of the newly generated sample type.

[0037] (4) When generating new sample operation data, the generation parameters are determined according to the distance from the support sample operation data, and the newly generated samples are located around the boundary, ensuring the reliability of the classification boundary. Description of the Drawings

[0038] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0039] Figure 1 It is a flowchart of the fault diagnosis method for a wind turbine based on SCADA data, system and electronic device according to an embodiment of the present invention;

[0040] Figure 2 It is a flowchart of the method for determining the generation parameters of the fault diagnosis method for a wind turbine based on SCADA data, system and electronic device according to an embodiment of the present invention;

[0041] Figure 3 It is a schematic diagram of the generation of new sample operation data in the fault diagnosis method for a wind turbine based on SCADA data, system and electronic device according to an embodiment of the present invention;

[0042] Figure 4 It is a flowchart of the method for determining the generation parameters in the fault diagnosis method for a wind turbine based on SCADA data, system and electronic device according to another embodiment of the present invention;

[0043] Figure 5 It is a flowchart of the operation of the system of the fault diagnosis method for a wind turbine based on SCADA data, system and electronic device according to an embodiment of the present invention;

[0044] Figure 6Schematic diagram of an electronic device according to an embodiment of the present invention. Detailed implementation manners

[0045] Next, in combination with the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0046] As Figure 1 shown, an embodiment of the present invention provides a wind turbine fault diagnosis method based on SCADA data, including the following steps:

[0047] S10: Obtain a plurality of sample operation data.

[0048] In this step, the sample operation data is the historical operation data of the wind turbine. The historical operation data is the recorded data of the wind turbine operation before model training. In the case of no recorded data, the wind turbine monitoring system can be sampled. When the wind turbine is running, the monitoring system collects the operation parameters of the wind turbine at a certain time interval and stores them for analyzing the operation state of the wind turbine.

[0049] In the embodiment of the present application, the wind turbine operation data includes the generator set related parameters, operation state parameters and fault state marks of the wind turbine.

[0050] Specifically, the generator set related parameters include the rated power of the wind turbine generator set, cut-in wind speed, cut-out wind speed, blade rotation speed, number of stages and total ratio of the transmission, rotation speed range of the generator, and rated rotation speed range of the generator.

[0051] For each turbine, two separate sets of data need to be provided: SCADA data and status data. Specifically, the SCADA data contains 58 parameters of each turbine at one-minute intervals. The status data provides information about the fault status. The relevant parameters of each wind turbine generator set in the wind power plant include rated power, Cut-in speed, Cut-out speed, rotation speed of the blade, number of stages and total ratio of the transmission, rotation speed range of the generator, and rated rotation speed range of the generator.

[0052] The SCADA data can be collected once per second. For each turbine, generally 58 parameters can be obtained in the SCADA data, and these parameters are divided into four categories.

[0053] 1) Vibration parameters: including vibration data such as radial vibration of the main shaft, axial vibration of the main shaft, and torsional vibration of the main shaft, which can reflect a series of problems such as mechanical looseness, bearing or gear faults.

[0054] 2) Health parameters: including main bearing temperature, low-speed shaft temperature, high-speed shaft temperature, and gearbox oil temperature, which are helpful for analyzing the health status of the wind turbine generator set.

[0055] 3) Performance parameters: including rotor speed, impeller speed, generator speed, active power, etc., which are used to measure the operating performance of the wind turbine generator set.

[0056] 4) Control parameters: referring to the status of the programmable logic controller code (PLC). There are 17 PLC status codes, namely system initialization, system interruption, system maintenance, system power generation, etc.

[0057] The status data provides information about the fault status, mainly including five parameters: wind turbine number, fault cause, maintenance activities for the fault, start date of maintenance, and end date of maintenance.

[0058] The fault status flag is the identification data of the fan fault status. A set of fan operation data corresponds to one status identifier, and the status identifiers for normal fan operation and fan fault are different.

[0059] For the obtained multiple sample operation data, they can be divided into positive sample operation data and negative sample operation data according to the fault status identifier among them. The fan fault status corresponding to the positive sample operation data is normal, and the fan fault status corresponding to the negative sample operation data is a fault.

[0060] In some embodiments, the obtained sample operation data is processed from the obtained historical monitoring data. Specifically, the directly obtained original operation data may be incorrect due to the failure of the data acquisition system. The obtained original operation data is cleaned to ensure the reliability of the obtained sample operation data. Specifically, errors are identified according to the attribute fields and record types in the obtained original operation data. In the range of attributes, there are two problems: missing values and incorrect field values. Missing values are generally values that are unavailable during data input, and incorrect field values are generally values with inappropriate data input positions. In the range of record types, there are two problems: duplicate records and conflicting records. Duplicate records are generally because the same record is represented twice; conflicting records are generally caused by records with the same key being described by different values. The original operation data with errors is deleted to clean the obtained original operation data.

[0061] In some embodiments, the amount of data after cleaning decreases. To expand the amount of data, the fan monitoring system is resampled to obtain sample operation data. For the resampled data, the data integrity is verified through the following formula:

[0062]

[0063] Where P di is data integrity; N r is the number of instances measured in a sampling period; N cc is data that meets consistency, that is, data that has passed the internal consistency check. Whether there is logical consistency between the various attributes in the internal consistency check data is checked here, and here the focus is on checking whether the values of the data are within a reasonable range. When the data integrity of the collected data reaches the preset value, the collected data is valid, otherwise resampling is performed. For example, resample the data at 10-minute intervals. If the proportion of data that meets consistency reaches 90%, the collected data is valid. The resampled SCADA data is defined as the data set of this method.

[0064] Most wind turbines are fault-free most of the time. Therefore, the number of normal instances in the data set is significantly more than the number of fault instances, which leads to the imbalance of the data set, so that the algorithm or model cannot accurately identify the complete signal characteristics, resulting in overfitting in the learning process and low reliability and accuracy of the results. Therefore, it is necessary to generate a balanced data set.

[0065] S20: Train at least part of the sample run data in the multiple sample run data through the support vector machine algorithm to determine the support sample run data.

[0066] The support vector machine (Support Vector Machine, SVM) algorithm is a supervised learning algorithm. Its core idea is to find an optimal hyperplane in the feature space for classification and make the interval between the two categories the largest.

[0067] In this step, at least part of the sample run data used to train the support vector machine includes positive sample run data and negative sample run data. The support sample run data is the boundary sample run data that distinguishes the positive sample run data and the negative sample run data. The distance from each support sample run data to the optimal hyperplane is the shortest. When training the support vector machine, any kernel function can be selected for training, such as the linear kernel function, the Gaussian kernel function, etc., and the embodiments of the present invention are not limited thereto.

[0068] S30: Determine the generation parameters according to the distance between the support sample run data and the K nearest neighbor sample run data.

[0069] In this step, generation parameters are generated as the calculation coefficients when generating new sample operation data. The generation parameters characterize the distance between the new samples and the operation data of the support samples. The larger the generation parameters, the closer the operation data of the new samples is to the operation data of the support samples. When generating multiple sets of operation data for new samples, determine the K nearest neighbor sample operation data that is closest to the operation data of each support sample. Here, K is an integer greater than 1, and the value of K can be selected according to the number of new sample operation data to be generated. The more new sample operation data to be generated, the larger the value of K; conversely, the fewer new sample operation data to be generated, the smaller the value of K.

[0070] For different support sample operation data, the number of selected nearest neighbor sample operation data can be the same or different. For any support sample operation data, with this support sample operation data as the center, calculate the distance between the surrounding sample operation data and it, and select the first K sample operation data according to the ascending order of the distances. For any one of the K sample operation data, determine the generation parameters according to the distance between it and this support sample operation data. The larger the distance, the larger the generation parameters. The embodiments of the present invention do not limit the specific method for determining the generation parameters. For example, in one implementation, K generation parameters with different magnitudes can be preset. When determining the generation parameters, select the corresponding generation parameters according to the distances between the K sample operation data and the same support sample operation data. Assume that K is taken as 3, and 3 generation parameters with different magnitudes can be preset, such as 0.2, 0.5, and 0.8. Among the 3 sample operation data, the generation parameter corresponding to the sample operation data closest to the support sample operation data is 0.8, and the generation parameter corresponding to the sample operation data second closest to the support sample operation data is 0.2.

[0071] S40: Generate multiple sets of new sample operation data according to the generation parameters to obtain the training data of the fan fault diagnosis model.

[0072] In this step, the position of the new sample operation data is between the support sample operation data and the nearest neighbor sample operation data. Calculate the distance difference between each nearest neighbor sample operation data and the support sample operation data. The sum of the product of this difference and the generation parameters and the support sample operation data is the new sample operation data. The new sample operation data is combined with the sample operation data in step S10 to obtain the training data of the fan fault diagnosis model.

[0073] In some embodiments, multiple new sample run data can be generated between a support sample run data and a nearest neighbor sample run data. That is, after determining the new sample run data according to the generation parameters, new sample run data can be regenerated between the new sample run data and the support sample run data, and new sample run data can also be regenerated between the new sample run data and the nearest neighbor sample data. The specific generation method is the same as that of steps S30 and S40. For the sake of brevity of description, it will not be elaborated here.

[0074] S50: Perform feature fusion on the training data to obtain fused training data.

[0075] Since not all features in the SCADA data are related to generator faults, first, a subset of features reflecting the operating state of the generator collected by the SCADA system is determined. More specifically, the analysis of the features of wind turbine faults belongs to the vibration parameters, health parameters, and performance parameters mentioned above. Table 1 summarizes the selected features for generator fault prediction and diagnosis purposes and their potential correlations.

[0076] Table 1 Inference rules for fault diagnosis

[0077]

[0078] The features described in Table 1 contain a certain degree of redundancy. The training data all include feature data of multiple dimensions, and each dimension of feature data corresponds to a parameter of the wind turbine, and there is a correlation between the features of each dimension. For example, the radial vibration data of the main shaft and the temperature of the main bearing are health data. The greater the vibration amplitude of the wind turbine, the greater the radial vibration parameter of the main shaft and the higher the temperature of the main bearing. The temperature of the low-speed shaft is closely related to the temperature of the high-speed shaft, and their correlation coefficient is greater than 0.98. The active power and reactive power of the generator are also closely related, and their correlation coefficient is greater than 0.95. Feature fusion is performed on the feature data of multiple dimensions, that is, multiple correlated feature data are represented by a fused feature to obtain fused training data. The specific fusion method can be the principal component analysis method, the linear discriminant analysis method, etc. The embodiments of the present invention are not limited to the specific fusion method. The dimension of the fused training data is significantly smaller than the dimension of the training data, thereby reducing the model training complexity.

[0079] S60: Train the fused training data through the support vector machine algorithm to obtain a wind turbine fault diagnosis model.

[0080] To monitor the operating state of the generator using the SCADA system, it is necessary to map from the feature space to the state space and determine whether the generator is in a normal operating state through features. This application realizes the ultimate goal of fault diagnosis of wind turbine units through the improved support vector machine algorithm SVM.

[0081] In this step, the improved support vector machine algorithm SVM uses the radial basis function RBF as the kernel function to map the fused training data into a high-dimensional space, so that the fused training data is linearly separable in the high-dimensional space. In some embodiments, other types of kernel functions, such as the Gaussian kernel function, etc., can also be used as the kernel function, and the embodiments of the present invention are not limited thereto. In the high-dimensional space, the model is prone to overfitting the training data, resulting in poor generalization ability on new data. Therefore, regularization is further added to the objective function to control the complexity of the model, prevent overfitting, and improve the generalization ability of the model.

[0082] The core of the improved support vector machine algorithm SVM is to find a hyperplane that maximizes the classification margin, and the standard Hinge loss can effectively punish misclassified samples and samples close to the classification margin. By minimizing the Hinge loss function, it helps to find a better classification hyperplane to better distinguish data of different classes.

[0083] The wind turbine fault diagnosis method based on SCADA data provided by the embodiments of the present invention can obtain a detection model by training historical operation data, which solves the problem of complex establishment of the detection model in the existing detection methods; new sample operation data is generated based on the obtained sample operation data during the model training process, increasing the number of samples and making the trained fault detection model more reliable; when generating new sample operation data, the boundary between positive sample operation data and negative sample operation data is determined by the support vector machine method, and new sample operation data is generated based on the support sample operation data on the boundary, ensuring the reliability of the newly generated sample types. When generating new sample operation data, the generation parameters are determined according to the distance from the support sample operation data, and the newly generated samples are located around the boundary, ensuring the reliability of the classification boundary.

[0084] In some embodiments, the number of positive sample operation data among multiple sample operation data is less than the number of negative sample operation data. In this case, positive class new sample operation data is generated to achieve the balance between positive class sample operation data and negative class sample operation data. Specifically, please refer to Figure 2 , Figure 2 which shows the method for determining the generation parameters in this embodiment, specifically including the following steps:

[0085] S301: Determine the positive support sample operation data in the support sample operation data.

[0086] In this step, the support sample operation data corresponding to the normal wind turbine fault state in the support sample operation data is the positive support sample operation data.

[0087] S302: Calculate the first distance between each positive support sample operation data and each positive sample operation data among the corresponding K nearest neighbor sample operation data.

[0088] In this step, the K nearest neighbor sample run data of a positive support sample run data includes positive sample run data and negative sample run data. The distance between a positive support sample run data and any positive sample run data among its corresponding K nearest neighbor sample run data is the first distance. The first distance can be the absolute value of the difference between the positive support sample run data and the nearest positive sample run data, or the Euclidean distance between the positive support sample run data and the nearest positive sample run data.

[0089] S303: Determine the generation parameters corresponding to each positive support sample run data according to the first distance.

[0090] In this step, the generation parameter reflects the distance between the newly generated sample and the positive support sample run data. For the same positive support sample run data, its corresponding generation parameter is the same, and the generation parameters corresponding to different positive support vectors are different. When multiple positive sample run data among the K nearest neighbor sample run data are close to the positive support sample run data, the generation parameter is small, and the newly generated sample run data is close to the nearest positive sample run data. When multiple positive sample run data among the K nearest neighbor sample run data are far from the positive support sample run data, the generation parameter is large, and the newly generated sample run data is close to the nearest positive support sample run data. Through the generation parameter, the decision boundary can be balanced while ensuring the category of the newly generated sample. In one embodiment, the average value of the first distances of each positive sample run data between a positive support sample run data and its corresponding K nearest neighbor sample run data is used as the generation parameter. In other embodiments, other methods can also be used, for example, using a pre-set distance and generation parameter correspondence table to determine the generation parameter.

[0091] After determining the generation parameter, generate new sample run data according to the method described in step S40 to obtain training data.

[0092] In some embodiments, when the number of positive sample run data among the multiple sample run data obtained is less than the number of negative sample run data, the newly generated sample run data is positive sample run data. The specific generation method is as follows:

[0093] When the number of positive sample run data among the K nearest neighbor sample run data corresponding to the positive support sample run data is less than the number of negative sample run data, the new sample run data is:

[0094] X new =x SV+ +δ(x nn+ -x SV+ )+αγ 0 ;

[0095] If the number of positive sample running data among the K nearest neighbor sample running data corresponding to the positive support sample running data is more than the number of negative sample running data, the new sample running data is:

[0096] X new = x SV+ - δ(x nn+ - x SV+ ) + αγ 0 ;

[0097] where x SV+ is any positive support sample running data, δ is a random number in the interval (0, 1), x nn+ is any positive sample running data among the K nearest neighbor sample running data corresponding to x SV+ , α is a random number in the interval (0, 1), and γ 0 is a generation parameter. Both δ and α are random numbers in the interval (0, 1), and they may be the same or different.

[0098] The generation parameter γ here 0 samples the first distance mean, and its expression is where k + is the number of positive class nearest neighbors; x SV is the minority class support vector; represents the j-th positive class nearest neighbor of x SV .

[0099] To more clearly illustrate the new sample generation method in the embodiments of the present invention, Figure 3 a schematic diagram of the new sample generation method is shown. As Figure 3 shown, the squares are positive sample running data, the circles are negative sample running data, and the number of positive sample running data is less than the number of negative sample running data. To achieve sample balance, positive sample running data needs to be generated. Assume that x i is any positive support sample running data among the support vectors, and the number of positive support sample running data among the K nearest neighbor sample running data corresponding to x i is more than the number of negative support sample running data. Any one of the positive sample running data is x k , then according to the above calculation method, the generated new sample running data is shown as triangles in the figure. x j is any positive support sample running data among the support vectors, and the number of negative support sample running data among the K nearest neighbor sample running data corresponding to x j is more than the number of positive support sample running data. Any one of the positive sample running data is x l , then according to the above calculation method, the generated new sample running data is shown as stars in the figure.

[0100] In this embodiment, when the number of positive sample operation data among the obtained multiple sample operation data is less than the number of negative sample operation data, the generated new sample operation data is positive sample operation data, which ensures the balance of the samples and makes the trained fault diagnosis model more reliable.

[0101] Further, if the positive sample operation data accounts for the majority among the nearest neighbor sample operation data corresponding to the positive support sample operation data, it can be known that this boundary sample is in the area where the two types of data are far apart. Therefore, extrapolation is performed on this positive support sample operation data; when the negative sample operation data accounts for the majority among the nearest neighbor sample operation data, it can be known that this positive support sample operation data sample is in the overlapping area of the positive and negative sample operation data in the kernel space. Therefore, interpolation is performed on this positive support sample operation data; in this way, the classification surface is made more ideal.

[0102] In addition, when generating new sample operation data, the generation parameter is determined according to the distance from the support sample operation data. Therefore, adding the generation parameter operation to generate new sample operation data can ensure that the generated new sample operation data is located around the boundary, ensuring the reliability of the classification boundary.

[0103] In some embodiments, the number of negative sample operation data among the multiple sample operation data is less than the number of positive sample operation data. In this case, negative class new sample operation data is generated to achieve the balance between the positive class sample operation data and the negative class sample operation data. Specifically, please refer to Figure 4 , Figure 4 which shows the method for determining the generation parameter in this embodiment, specifically including the following steps:

[0104] S311: Determine the negative support sample operation data in the support sample operation data.

[0105] In this step, the support sample operation data corresponding to the fan fault state calibrated as a fault in the support sample operation data is the negative support sample operation data.

[0106] S312: Calculate the third distance between each negative support sample operation data and each negative sample operation data in its corresponding K nearest neighbor sample operation data.

[0107] In this step, the K nearest neighbor sample operation data of a negative support sample operation data includes positive sample operation data and negative sample operation data. The distance between a negative support sample operation data and any positive sample operation data in its corresponding K nearest neighbor sample operation data is the third distance. The third distance can be the absolute value of the difference between the negative support sample operation data and the nearest neighbor negative sample operation data, or the Euclidean distance between the negative support sample operation data and the nearest neighbor negative sample operation data.

[0108] S313: Determine the generation parameters corresponding to each negative support sample operation data according to the third distance.

[0109] In this step, the generation parameters reflect the distance between the newly generated samples and the negative support sample operation data. For the same negative support sample operation data, the corresponding generation parameters are the same, and different negative support vectors have different corresponding generation parameters. When multiple negative sample operation data among the K nearest neighbor sample operation data are close to the negative support sample operation data, the generation parameters are small, and the newly generated sample operation data is close to the nearest neighbor negative sample operation data. When multiple negative sample operation data among the K nearest neighbor sample operation data are far from the negative support sample operation data, the generation parameters are large, and the newly generated sample operation data is close to the nearest neighbor negative support sample operation data. Through the generation parameters, the decision boundary can be balanced while ensuring the category of the newly generated samples. In one embodiment, the average value of the third distances between a negative support sample operation data and each negative sample operation data among its corresponding K nearest neighbor sample operation data is used as the generation parameter. In other embodiments, other methods can also be adopted. For example, a pre-set distance-generation parameter correspondence table can be used to determine the generation parameters.

[0110] After determining the generation parameters, generate new sample operation data according to the method described in step S40 to obtain training data.

[0111] In some embodiments, when the number of negative sample operation data among the multiple obtained sample operation data is less than the number of positive sample operation data, new sample operation data is generated through the following method:

[0112] When the number of negative sample operation data among the K nearest neighbor sample operation data corresponding to the negative support sample operation data is less than the number of positive sample operation data, the new sample operation data is:

[0113] X new =x SV- +δ(x nn- -x SV- )+βγ 0 ;

[0114] When the number of negative sample operation data among the K nearest neighbor sample operation data corresponding to the negative support sample operation data is more than the number of positive sample operation data, the new sample operation data is:

[0115] X new =x SV- -δ(x nn- -x SV- )+βγ 0 ;

[0116] where x SV-Run data for any negative support sample, δ is a random number in the interval (0, 1), x nn - Run data for x SV- Any negative sample run data among the run data of the K nearest neighbor samples corresponding to x, β is a random number in the interval (0, 1), γ 0 Is a generation parameter. Both δ and β are random numbers in the interval (0, 1), and they may be the same or different.

[0117] The generation parameter γ here 0 Selects the third distance mean, and its expression is Where Represents the j-th negative class nearest neighbor of x SV , k - Is the number of negative class nearest neighbors.

[0118] It should be understood that δ in this embodiment and δ in the embodiment of generating positive sample run data are both random numbers, and they may be the same or different.

[0119] In this embodiment, when the number of negative sample run data in the obtained multiple sample run data is less than the number of positive sample run data, the generated new sample run data is negative sample run data, which ensures the balance of the samples and makes the trained fault diagnosis model more reliable.

[0120] Furthermore, if the negative sample run data accounts for the majority among the nearest neighbor sample run data corresponding to the negative support sample run data, it can be known that the negative sample run data is in the area where the two types of data are far apart, so extrapolation is performed on the negative support sample run data; if the positive sample run data accounts for the majority among the nearest neighbor sample run data corresponding to the negative support sample run data, it can be known that the boundary sample is in the overlapping area of the positive sample run data and the negative sample run data in the kernel space, so interpolation is performed on the negative support sample run data; in this way, the classification surface is made more ideal. In addition, when generating new sample run data, the generation parameter is determined according to the distance from the support sample run data. Therefore, adding the generation parameter operation to generate new sample run data can ensure that the generated new sample run data is located around the boundary, ensuring the reliability of the classification boundary.

[0121] For step S50, in some embodiments, an objective function is constructed by local linear embedding and principal component analysis to perform feature fusion on the training data. Specifically, a neighborhood set corresponding to each training data is determined, and the neighborhood set includes N training data that are closest to the corresponding training data; the linear weight coefficients of each training data and the N training data in the corresponding neighborhood set are calculated; the preset objective function is optimized based on the linear weight coefficients until the preset objective function converges to obtain fused training data; wherein the preset objective function includes regularization and the sum of errors. The error includes the error corresponding to the local linear embedding algorithm and the error corresponding to the principal component analysis method.

[0122] Specifically, for a high-dimensional dataset X∈R n×d For each training data x i , n represents the number of samples in the high-dimensional dataset, d represents the number of features, and calculates its difference from all other training data x j The Euclidean distance of where x i and x j Respectively represent the i-th and j-th training data of the high-dimensional data set, x ij represents the kth feature of the i-th training data, x jk Represents the kth feature of the jth training data; for each training data x i Select the N training data with the closest Euclidean distance as the neighborhood set N(x i ).

[0123] For each training data x i , which corresponds to any training data x in the N training data in the neighborhood set j The linear weight coefficient is w ij : in, Keeping the weight composition unchanged, solve the low-dimensional representation Y: This part ensures a low-dimensional representation of each data point y in Y i It can be reconstructed by a linear combination of its domain, thus preserving the local structure. Then the objective function J(Y) is optimized using gradient descent.

[0124] The preset objective function is optimized using the gradient descent method based on the linear weight coefficient as J(Y):

[0125]

[0126] Among them, y i and j The training data x i and x jThe corresponding fault status flag. The linear weight coefficient regularization term prevents overfitting and ensures the sparsity of the weights. Trace means to find the trace of a two-dimensional square matrix. Through iterative update by the gradient descent method until the objective function converges to a stable value, the fused training data after dimensionality reduction is finally output. Regularization term Prevents overfitting and ensures the sparsity of the weights. λ is the regularization parameter. From the gradient calculation formula The gradient can be obtained where I is the identity matrix and W i is the weight matrix represents the local structure. Since the derivative of the regularization term with respect to y i does not directly affect y i , the gradient is 0. 2αY represents the global structure. αj is the weight parameter used to balance the importance of the local structure and the global structure. Updating Y in the direction of the gradient in this way, the updated low-dimensional representation is obtained: where η is the learning rate. Iteratively update the low-dimensional representation Y until the objective function J(Y) converges to a stable value. The finally output data Y after dimensionality reduction

[0127] During the dimensionality reduction process, the local structure of the original data is retained, so that the low-dimensional representation after dimensionality reduction can still reflect the local neighborhood relationship of the original data. This method of retaining the local structure makes the result after dimensionality reduction have better interpretability, can more clearly reflect the local features and mutual relationships of the original data, and is suitable for subsequent classification tasks

[0128] For step S60, when training the fan fault diagnosis model, the improved support vector machine algorithm SVM selects the RBF kernel function and uses the standard Hinge loss plus regularization as the objective function to optimize the parameters to better solve the problems of non-linear feature mapping and high-dimensional data. The specific objective function is expressed as:

[0129]

[0130] where y i and y j are the corresponding fault status flags of the training data x i and x j respectively; w is the model weight vector; b is the bias constant is the regularization term; C is the regularization constant that controls the trade-off between the Hinge loss and the regularization term; γ is the hyperparameter of the RBF kernel function and is a preset constant. The objective function is used to train the model, and w and b are optimized during the training process. α j is the support vector

[0131] After obtaining the optimized \(w\) and \(b\) through training, instead of directly using the parameter \(w\) in the decision function, the kernel function is utilized to calculate the dot product without explicitly computing the high-dimensional mapping, and the optimized parameters are directly used to quickly classify the new sample \(x\). The decision function is expressed as:

[0132] \(\alpha\) i is the support vector. The improved support vector machine algorithm SVM can effectively process high-dimensional data and perform non-linear classification without explicitly computing the high-dimensional mapping, and thus obtain the label indicating whether the wind turbine has a fault. Combining with Table 1, the location and cause of the wind turbine fault can be determined.

[0133] Figure 5 A wind turbine fault diagnosis method according to an embodiment of the present invention is shown, which specifically includes the following steps:

[0134] S100: Obtain the wind turbine operation data.

[0135] S200: Perform feature fusion on the wind turbine operation data to obtain the fused data.

[0136] S300: Input the fused data into the wind turbine fault diagnosis model to diagnose the wind turbine fault.

[0137] In this embodiment, the obtained wind turbine operation data has the same dimension as the sample operation data obtained during the training of the wind turbine fault diagnosis model, and the features represented by each dimension also correspond one by one. The method of performing feature fusion is the same as the method of performing feature fusion on the training data in step S50. Please refer to the training feature fusion method in step S50, which will not be elaborated here. The wind turbine fault diagnosis model used in this embodiment is the model obtained by the wind turbine fault diagnosis method based on SCADA data in any of the above embodiments.

[0138] Through this embodiment, the wind turbine fault can be diagnosed only by inputting the obtained wind turbine operation data into the wind turbine fault diagnosis model, which greatly reduces the cost and complexity of the wind turbine fault diagnosis and improves the efficiency of the wind turbine fault diagnosis.

[0139] Figure 6 A schematic diagram of an electronic device according to an embodiment of the present invention is shown. As Figure 6 shown, the electronic device according to an embodiment of the present invention includes a memory 61 and a processor 62. The memory 61 stores a calculation program, which can be executed by the processor 62. When the calculation program is executed by the processor 62, the processor 62 can implement the steps corresponding to the wind turbine fault diagnosis method based on SCADA data in any of the above embodiments, and / or the processor 62 can implement the steps corresponding to the wind turbine fault diagnosis method in any of the above embodiments.

[0140] It should be understood that in the embodiments of the present invention, the memory 61 can be any kind of computer-readable memory, such as RAM, Flash, etc. The embodiments of the present invention are not limited to the specific type of the memory 61. The processor 62 can be one or more, and the processor 62 can execute the logic corresponding to the calculation program and output the calculation result.

[0141] The embodiments of the present invention also provide a computer-readable storage medium, in which a calculation program is stored. When the calculation program is executed, the steps corresponding to the wind turbine fault diagnosis method based on SCADA data in any one of the above embodiments can be implemented, and / or the steps corresponding to the wind turbine fault diagnosis method in any one of the above embodiments can be implemented.

[0142] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A wind turbine fault diagnosis method based on SCADA data, characterized in that: The method comprises: Acquire a plurality of sample operation data, wherein the sample operation data refers to operation data corresponding to a fan fault state that has been calibrated, and the plurality of sample operation data includes a plurality of positive sample operation data and a plurality of negative sample operation data, wherein the fan fault state corresponding to the positive sample operation data is normal, and the fan fault state corresponding to the negative sample operation data is fault; Training at least part of the sample operation data among the plurality of sample operation data by a support vector machine algorithm to determine support sample operation data, wherein the support sample operation data is boundary sample operation data for distinguishing positive sample operation data from negative sample operation data; Determine the generation parameter according to the distance between the supporting sample running data and the K nearest neighbor sample running data, where K is an integer greater than 1; Generate a plurality of new sample operation data according to the generation parameters to obtain training data of a fan fault diagnosis model, wherein the training data includes the plurality of sample operation data and the plurality of new sample operation data; Performing feature fusion on the training data to obtain fused training data; The fused training data is trained by a support vector machine algorithm to obtain a fan fault diagnosis model; When the number of positive sample running data in the plurality of sample running data is less than the number of negative sample running data, determining the generation parameter according to the distance between the supporting sample running data and the K nearest neighboring sample running data includes: determining positive supporting sample operation data among the supporting sample operation data; Calculate the first distance between each positive support sample running data and each positive sample running data in the corresponding K nearest neighbor sample running data; Determine, according to the first distance, a generation parameter corresponding to each of the positive supporting sample running data; At this time, a plurality of new sample operation data are generated according to the generation parameters, including: When the number of positive sample running data in the K nearest neighbor sample running data corresponding to the positive support sample running data is less than the number of negative sample running data, the new sample running data is: X new =x SV+ +δ(x nn+ -x SV+ )+αγ0; When the number of positive sample running data in the K nearest neighbor sample running data corresponding to the positive support sample running data is greater than the number of negative sample running data, the new sample running data is: X new =x SV+ -δ(x nn+ -x SV+ )+αγ0; Among them, x SV+ is any one of the positive support sample running data, δ is a random number in the interval (0, 1), x nn+ For x SV+ Any positive sample running data among the corresponding K nearest neighbor sample running data, α is a random number in the interval (0, 1), and γ0 is the generation parameter; When the number of positive sample running data in the plurality of sample running data is greater than the number of negative sample running data, determining the generation parameter according to the distance between the supporting sample running data and the K nearest neighboring sample running data includes: determining negative supporting sample running data among the supporting sample running data; Calculate the third distance between each negative support sample running data and each negative sample running data in the corresponding K nearest neighbor sample running data; Determine, according to the third distance, a generation parameter corresponding to each of the negative support sample running data; At this time, a plurality of new sample operation data are generated according to the generation parameters, including: When the number of negative sample running data in the K nearest neighbor sample running data corresponding to the negative support sample running data is less than the number of positive sample running data, the new sample running data is: X new =x SV- +δ(x nn- -x SV- )+βγ0; When the number of negative sample running data in the K nearest neighbor sample running data corresponding to the negative support sample running data is greater than the number of positive sample running data, the new sample running data is: X new =x SV- -δ(x nn- -x SV- )+βγ0; Among them, x SV- Run the data for any of the negative support samples, δ is a random number in the interval (0,1), x nn- For x SV- For any negative sample running data among the corresponding K nearest neighbor sample running data, β is a random number in the interval (0,1).

2. The wind turbine fault diagnosis method based on SCADA data as claimed in claim 1, characterized in that: The determining, according to the first distance, a generation parameter corresponding to each positive supporting sample running data includes: The mean of the first distances is calculated to obtain the generation parameter.

3. The wind turbine fault diagnosis method based on SCADA data as claimed in claim 1, characterized in that: The determining, according to the third distance, a generation parameter corresponding to each of the negative support sample running data comprises: The mean of the third distance is calculated to obtain the generation parameter.

4. The wind turbine fault diagnosis method based on SCADA data according to claim 1 is characterized in that: The step of performing feature fusion on the training data to obtain fused training data includes: Determine a neighborhood set corresponding to each of the training data, wherein the neighborhood set includes N training data that are closest to the corresponding training data; Calculate the linear weight coefficient of each training data and N training data in the corresponding neighborhood set; The preset objective function is optimized based on the linear weight coefficient until the preset objective function converges to obtain the fusion training data; the preset objective function includes regularization and the sum of errors.

5. A wind turbine fault diagnosis system based on SCADA data, characterized in that: The system comprises: A data acquisition module, used to acquire fan operation data; A feature fusion module performs feature fusion on the wind turbine operation data to obtain fused data; A fault diagnosis module inputs the fused data into a wind turbine fault diagnosis model to perform fault diagnosis on the wind turbine; wherein the wind turbine fault diagnosis model is obtained according to the wind turbine fault diagnosis method based on SCADA data as described in any one of claims 1 to 4.

6. An electronic device, characterized in that: The electronic device includes a memory and a processor, the memory stores a calculation program, and when the calculation program is executed by the processor, the processor implements the wind turbine fault diagnosis method based on SCADA data as described in any one of claims 1-4.

Images