Wind farm-level wind turbine blade fault detection method, device and storage medium

Abstract

The present invention relates to the field of neural network technology, and in particular to a wind farm-level wind turbine blade fault detection method, the method comprising: obtaining a target detection model of a wind farm-level wind turbine and a test data set of each wind turbine in a plurality of wind turbines, the target detection model comprising: N BiLSTM layers, N-1 dropout layers and a fully connected layer, two adjacent BiLSTM layers are connected by a dropout layer, and the output end of the last BiLSTM layer is connected to the fully connected layer; preprocessing the test data set of each wind turbine to obtain a plurality of sub-data sets; normalizing the plurality of sub-data sets to obtain a plurality of target sub-data sets; detecting and processing the plurality of target sub-data sets by a target detection model to obtain blade fault type information of each wind turbine. The method improves the detection precision, accuracy and generalization ability of wind turbines in a wide range of wind farms.

Description

Technical Field

[0001] The present invention relates to the technical field of wind turbine blade fault detection based on data identification, and in particular to a wind farm-level wind turbine blade fault detection method, device and storage medium. Background Art

[0002] Wind turbines are often located at high altitudes and offshore areas with abundant wind energy, which puts their components, especially blades, at a higher risk of failure. Blade failure is the most common type of accident in wind turbines. This highlights the critical need for effective monitoring, timely fault detection and diagnosis of wind turbine blades to prevent serious blade failures and reduce wind turbine downtime, thereby ensuring the efficiency and reliability of turbines in harsh environments.

[0003] Existing wind turbine detection methods mainly use machine learning techniques such as decision trees, support vector machines, and random forests to analyze wind turbine data, which can detect subtle patterns and correlations of potential faults. However, most of the models proposed by existing detection methods are verified on a single wind turbine. Models trained on data from a single turbine are often over-adapted to the specific characteristics and operating mode of that turbine, resulting in a lack of universality and accuracy when applied to different turbines or under different conditions. Large wind farms are usually composed of dozens to hundreds of turbines. Applying existing detection methods to large wind farms will cause the problem of low detection accuracy of wind turbines in these huge wind farms. Summary of the invention

[0004] The embodiments of the present application provide a wind farm-level wind turbine blade fault detection method, device and storage medium, which solves the technical problem that the existing detection method is applied to large wind farms, which will cause low detection accuracy of wind turbines in these huge wind farms. It achieves technical effects such as detecting blade fault types of wind turbines in a wide range of wind farms through multiple data of multiple wind turbines, improving the detection precision and accuracy of wind turbines in a wide range of wind farms, and improving the identification and detection efficiency of wind turbines in a wide range of wind farms.

[0005] In a first aspect, an embodiment of the present invention provides a wind farm-level wind turbine blade fault detection method, comprising: obtaining a target detection model of a wind farm-level wind turbine and a data set to be tested for each wind turbine in a plurality of wind turbines, wherein the target detection model comprises: N bidirectional long short-term memory network layers BiLSTM layers, N-1 random dropout layers dropout layers and a fully connected layer, N>1, two adjacent BiLSTM layers are connected by one dropout layer, and the output end of the last BiLSTM layer among the N BiLSTM layers is connected to the fully connected layer;

[0006] Preprocessing the data set to be tested of each wind turbine to obtain a plurality of sub-data sets;

[0007] Normalizing the multiple sub-datasets to obtain multiple target sub-datasets;

[0008] The target detection model is used to perform detection processing on the multiple target sub-datasets to obtain blade fault type information of each wind turbine.

[0009] Based on the same inventive concept, in a second aspect, the present invention provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of a wind farm-level wind turbine blade fault detection method are implemented.

[0010] Based on the same inventive concept, in a third aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the steps of a wind farm-level wind turbine blade fault detection method are implemented.

[0011] One or more technical solutions in the embodiments of the present invention have at least the following technical effects or advantages:

[0012] In an embodiment of the present invention, after obtaining a target detection model of a wind turbine at a wind farm level and a data set to be tested of each wind turbine in a plurality of wind turbines, the data sets to be tested of the plurality of wind turbines are first preprocessed to obtain a plurality of sub-data sets, so as to ensure the integrity of the data sets to be tested of the plurality of wind turbines, so as to facilitate the subsequent efficient and accurate detection by the target detection model. Then, the plurality of sub-data sets are normalized to obtain a plurality of target sub-data sets, so that the sub-data in each sub-data set are at a consistent scale, and a plurality of target sub-data sets are obtained, so as to further improve the efficient and accurate detection process of the target detection model for the data set to be tested, and also speed up the detection process of the target detection model. Then, the plurality of target sub-data sets are detected and processed by the target detection model to obtain the blade fault type information of each wind turbine. Since the target detection model has a multi-layer bidirectional long short-term memory network layer BiLSTM, the random discarding strategy of the model is enhanced, so that the target detection model avoids overfitting during the detection process, and improves the generalization ability of fault diagnosis of wind turbines at the wind farm level. In this way, the target sub-dataset is detected and processed through the target detection model, which improves the effectiveness of the detection process and the recognition detection efficiency, and enhances the detection precision and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present invention. Also, throughout the accompanying drawings, the same reference figures are used to represent the same components. In the drawings:

[0014] Figure 1 A schematic flow chart of the steps of a wind farm-level wind turbine blade fault detection method in an embodiment of the present invention is shown;

[0015] Figure 2 A schematic flow chart of the steps of performing iterative model training on an initial detection model through training sets of multiple wind turbines and performing detection on a data set to be tested of each wind turbine in the multiple wind turbines in an embodiment of the present invention is shown;

[0016] Figure 3 A schematic diagram showing the structure of a correlation matrix of first training data in an embodiment of the present invention is shown;

[0017] Figure 4 A schematic diagram showing the structure of a process of processing each target feature variable of target sub-training data by a single BiLSTM layer of a detection model in an embodiment of the present invention;

[0018] Figure 5 A structural schematic diagram showing the processing process of a single LSTM unit in an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0019] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0020] Embodiment 1

[0021] A first embodiment of the present invention provides a wind farm-level wind turbine blade fault detection method, such as Figure 1 As shown, including:

[0022] S101, obtaining a target detection model of a wind turbine at a wind farm level and a data set to be tested for each wind turbine in a plurality of wind turbines, wherein the target detection model includes: N bidirectional long short-term memory network layers BiLSTM layers, N-1 random dropout layers dropout layers and a fully connected layer, N>1, two adjacent BiLSTM layers are connected by a dropout layer, and an output end of the last BiLSTM layer among the N BiLSTM layers is connected to the fully connected layer;

[0023] S102, preprocessing the data set to be tested of each wind turbine to obtain multiple sub-data sets;

[0024] S103, performing normalization processing on the multiple sub-datasets to obtain multiple target sub-datasets;

[0025] S104, performing detection processing on the multiple target sub-datasets through the target detection model to obtain blade fault type information of each wind turbine.

[0026] It should be noted that BiLSTM stands for Bi-directional Long Short-Term Memory, which is called bidirectional long short-term memory network layer in Chinese. Multiple wind turbines are in the same large-scale wind farm. In a large-scale wind farm, the test data set of each wind turbine is the SCADA data to be tested of each wind turbine, that is, the SCADA data to be tested of each wind turbine is the data obtained through the SCADA (Supervisory Control And Data Acquisition) system.

[0027] After obtaining the target detection model of the wind turbine at the wind farm level and the data set to be tested of each wind turbine in the multiple wind turbines, the data sets to be tested of the multiple wind turbines are first preprocessed to obtain multiple sub-data sets to ensure the integrity of the data sets to be tested of the multiple wind turbines, so as to facilitate the subsequent efficient and accurate detection by the target detection model. Then, the multiple sub-data sets are normalized to obtain multiple target sub-data sets, so that the sub-data in each sub-data set are at a consistent scale, and multiple target sub-data sets are obtained, which further improves the efficient and accurate detection process of the target detection model for the data set to be tested (i.e., the target sub-data set), and can also speed up the detection process of the target detection model. Then, the multiple target sub-data sets are detected and processed by the target detection model to obtain the blade fault type information of each wind turbine. Since the target detection model has a multi-layer bidirectional long short-term memory network layer (Bi-directional LongShort-Term Memory, abbreviated as BiLSTM), the random discarding strategy of the model is enhanced, so that the target detection model avoids overfitting during the detection process, and improves the generalization ability of fault diagnosis of wind turbines at the wind farm level. In this way, the target sub-dataset is detected and processed through the target detection model, which improves the effectiveness of the detection process and the recognition detection efficiency, and enhances the detection precision and accuracy.

[0028] Next, combine Figure 1 The specific implementation steps of the wind farm-level wind turbine blade fault detection method provided by this embodiment are described in detail:

[0029] First, execute step S101 to obtain a target detection model of a wind turbine at a wind farm level and a data set to be tested for each wind turbine in a plurality of wind turbines, wherein the target detection model includes: N bidirectional long short-term memory network layers BiLSTM layers, N-1 random dropout layers and a fully connected layer, N>1, two adjacent BiLSTM layers are connected by a dropout layer, and the output end of the last BiLSTM layer among the N BiLSTM layers is connected to the fully connected layer.

[0030] Specifically, the data set to be tested of each wind turbine in the multiple wind turbines at the wind farm level is acquired through the SCADA system. The way to acquire the target detection model is: the initial detection model is iteratively trained through the training set of each wind turbine in the multiple wind turbines that have been acquired, until the current loss function value corresponding to the iterative detection model is not greater than the loss function threshold, and the current detection model is used as the target detection model. Among them, the current loss function value of the iterative detection model is obtained through the focal loss function. The training set of each wind turbine is a training data set acquired through the SCADA system. In this embodiment, the training set and the data set to be tested of the wind turbine are acquired through the SCADA system. The data signal of the SCADA system comes from the existing infrastructure in the wind turbine control system. The use of the data signal of the SCADA system is perfectly consistent with the goal of fault detection and diagnosis, providing a cost-effective and efficient continuous monitoring and maintenance solution. It should also be noted that in this embodiment, the model training process and principle of the initial detection model for the training set of the wind turbine are consistent with the detection process (i.e., reasoning process) and principle of the target detection model for the data set to be tested of the wind turbine.

[0031] The data format of the training set and the data set to be tested of each wind turbine is that, for each wind turbine, each training data or test data of the wind turbine is collected in the order of time to obtain the training set or test data set of the wind turbine. For example, the training data or test data of the wind turbine is collected once every preset time of 6 seconds, that is, one training data or test data of the wind turbine is collected every preset time of 6 seconds. The preset time can be set according to actual needs. Taking a certain wind turbine as an example to further illustrate, on a certain day, the first training data of the wind turbine is collected at 00:00:00, the second training data of the wind turbine is collected at 00:00:06, and the third training data of the wind turbine is collected at 00:00:12. The training data of the wind turbine is collected once every 6 seconds, and so on, to obtain the training set of the wind turbine on that day.

[0032] A training data or a test data of a wind turbine carries at least 2 characteristic variables. Optionally, in this embodiment, a training data or a test data of a wind turbine carries at least 27 characteristic variables. As shown in Table 1, the expression form of the 27 characteristic variables carried by a training data or a test data of a wind turbine. For example, in Table 1, the sequence number 1 of the characteristic variable represents the deflection angle of the first blade of the wind turbine, the sequence number 2 of the characteristic variable represents the deflection angle of the second blade of the wind turbine, the sequence number 3 of the characteristic variable represents the deflection angle of the third blade of the wind turbine, and so on. In Table 1, the sequence number 20 of the characteristic variable represents the active power of the wind turbine.

[0033] Table 1

[0034]

[0035] It should also be noted that the training data of the 27 characteristic variables carried by each wind turbine form a training set, and then a training set of multiple wind turbines is obtained. The prediction results corresponding to the training sets of multiple wind turbines include 19 different types of blade faults and 1 normal type, as shown in Table 2. In Table 2, the serial number 0 of the blade fault type indicates that the wind turbine is a normal type, and the normal type indicates that the wind turbine is in a normal operating state, indicating that the wind turbine has no blade fault. The serial number 1 of the blade fault type indicates that the wind turbine has a blade fault of the transverse crack type, the serial number 2 of the blade fault type indicates that the wind turbine has a blade fault of the trailing edge crack type, and so on.

[0036] Table 2

[0037]

[0038] This embodiment uses a large amount of SCADA data of multiple wind turbines for model training and model detection to improve the generalization ability of the detection model of this embodiment, and enhance the recognition detection efficiency and detection accuracy of the detection model. A large amount of SCADA data shows its extensiveness in both space and time. The model proposed using the SCADA data of wind turbines is effective in avoiding overfitting and significantly enhances its generalization ability. In addition, the SCADA data of wind turbines is used to perform blade fault detection on wind turbines to detect the blade fault types of wind turbines in a wide range of wind farms, improve the detection precision and accuracy of wind turbines in a wide range of wind farms, improve the recognition detection efficiency of wind turbines in a wide range of wind farms, and improve the effectiveness of the detection process.

[0039] like Figure 2 As shown, after obtaining the training set of each wind turbine in the multiple wind turbines, the process of iteratively training the initial detection model through the training sets of the multiple wind turbines is: step 1, preprocessing the training set of each wind turbine to obtain multiple sub-training sets. Step 2, normalizing the multiple sub-training sets to obtain multiple sub-target training sets. Step 3, iteratively training and detecting the multiple sub-target training sets through the initial detection model to obtain the blade fault prediction result of each wind turbine, and converging the model of the initial detection model through the focal loss function.

[0040] Step 1: preprocess the training set of each wind turbine to obtain multiple sub-training sets. The specific process is:

[0041] For each wind turbine training set, (1) missing value processing is performed on the wind turbine training set to obtain a first training data set, wherein the missing value processing is to remove the training data with missing values; (2) feature selection is performed on the first training data set to obtain a second training data set; (3) time series division is performed on the second training data set to obtain a sub-training set of the wind turbine, wherein the time series division is a process of dividing the second training data set according to a preset time step.

[0042] The above operations for obtaining a sub-data set of a wind turbine are performed for each training set of a wind turbine, that is, steps (1) to (3) are performed for each training set of a wind turbine to obtain multiple sub-training sets.

[0043] Specifically, for the training set of each wind turbine, step (1) performs missing value processing on the training set of the wind turbine to obtain the specific process of the first training data set: take the training set collected and obtained from a certain wind turbine on a certain day as shown in Table 3 as an example, and assume that the omitted part in Table 3 is filled with data. In Table 3, the first column represents the number of each training data in the training set of the wind turbine, the second column represents the sampling time corresponding to each training data, and the third to twenty-ninth columns represent the data of each characteristic variable carried by each training data. The characteristic variable 1 (i.e., the characteristic variable with the serial number 1) of the training data with training data number 1 (i.e., the first training data) represents the deflection angle of the first blade of the wind turbine of the first training data. It can be seen from Table 3 that the data of the characteristic variable 2 and the characteristic variable 3 of the second training data are both empty, indicating that the second training data is the training data with missing values, and the second training data is removed. Similarly, the data of the feature variable 2 of the third training data is empty, indicating that the third training data is training data with missing values, so the third training data is removed.

[0044] Table 3

[0045]

[0046] The training data shown in Table 4 are obtained by removing the second training data and the third training data in the training set of the wind turbine shown in Table 3. The training set after removing the training data with missing values ​​is the first training data set, and the training data shown in Table 4 is the obtained first training data set of the wind turbine. Optionally, the numbers of the training data in Table 4 can be renumbered or the original numbers can be used.

[0047] Table 4

[0048]

[0049] In this embodiment, the missing values ​​in the original SCADA data of the wind turbine are processed by simple and effective missing value processing, that is, the instances with incomplete data are deleted to ensure the data quality in the subsequent training process. At the same time, each first training data in the first training data set of each wind turbine carries the sampling time (i.e., the time interval feature), retains valuable time information, and provides a basis for the subsequent calculation of the interval between consecutive data points. In this way, while maintaining the integrity of the data, the key time dynamic information is also retained to improve the generalization ability of the model, and enhance the detection accuracy, precision and recognition detection efficiency of the model.

[0050] Since data containing time information is crucial to the accuracy and efficiency of the model, and each data carries too many feature variables, especially feature variables that have little impact on the model output, which will lead to unnecessary computational overhead for the model, it is necessary to screen the feature variables of the entire data set to avoid increasing the risk of overfitting. Therefore, step (2) is executed to perform feature selection on the first training data set to obtain the second training data set.

[0051] The specific process of step (2) is as follows: ①, perform correlation processing on any two characteristic variables of each first training data in the first training data set to obtain the correlation matrix of each first training data. ②, perform matrix averaging operation on the correlation matrix of each first training data to obtain the correlation matrix of the first training data set, and obtain the specified characteristic variable according to the correlation matrix of the first training data set. ③, remove the specified characteristic variable from the characteristic variables of each first training data to obtain each first training data after removal, and the characteristic variable of the first training data after removal is the characteristic variable of the first data after removal. ④, determine each second training data according to each first training data after removal, and obtain the second training data set according to each second training data.

[0052] In step (2), ①, for each first training data in the first training data set, any two feature variables of the first training data are subjected to correlation processing as shown in formula (1) to obtain a correlation matrix for each first training data. Formula (1) is as follows:

[0053] (1);

[0054] In the training phase, n1 represents the total number of feature variables in the first training data, t1 is the index of the feature variable of the first training data, x1 and x2 represent any two feature variables in the first training data, and Represent x1 and The mean of . It is obtained by performing the average operation on x1 of each first training data set. Similarly, we can also get . Represents the correlation matrix of the first training data.

[0055] In the inference phase, n1 represents the total number of feature variables in the first data, t1 is the index of the feature variable of the first data, x1 and x2 represent any two feature variables in the first data, and Represent x1 and The mean of . It is obtained by performing the average operation on x1 of each first data of the first data set. Similarly, we can also get . A correlation matrix representing the first data.

[0056] Each first training data performs the operation of step (2) to obtain a correlation matrix of each first training data, such as the correlation matrix of a first training data shown in Table 5.

[0057] Table 5

[0058]

[0059] After obtaining the correlation matrix of each first training data, execute ② in step (2). In ② of step (2), perform a matrix averaging operation on the correlation matrix of each first training data to obtain the correlation matrix of the first training data set, and obtain the specified characteristic variable based on the correlation matrix of the first training data set. Among them, the specified characteristic variable is a characteristic variable that has a weak correlation with characteristic variables other than its own characteristic variable, and the threshold of the weak correlation can be set according to actual needs. Specifically, each element in each first training data is added and averaged to obtain the corresponding element in the correlation matrix of the first training data set. Figure 3 The correlation matrix of the first training data is shown in Figure 3 In the figure, the data changes from dark gray to dark red, and different colors correspond to the element values ​​in the correlation matrix of different first training data. Figure 3 It can be seen that there is a weak correlation between feature variables 4 to 6 and most other feature variables, but feature variables 4 to 6 have a strong correlation with each other. Feature variables 22 to 25 also show similar correlation relationships. Therefore, feature variables 4-6 and 22-25 are determined as designated feature variables.

[0060] After the designated characteristic variables are determined, step (2) ③ is executed. In step (2) ③, the designated characteristic variables are removed from the characteristic variables of each first training data to obtain each first training data after removal, and the characteristic variables of the first training data after removal are the characteristic variables of the first data after removal. Specifically, after the designated characteristic variables are determined, for each first training data, the designated characteristic variables in the characteristic variables of the first training data are deleted to obtain the first training data after removal. Step ③ is executed for each training data to obtain each first training data after removal. For example, the characteristic variables 4-6 and 22-25 of the first first training data in the first training data set are deleted to obtain the first first training data after removal. The characteristic variables of the first first training data after removal are characteristic variables 1-3, 7-21, and 26-27. The characteristic variables 4-6 and 22-25 of the second first training data in the first training data set are deleted to obtain the second first training data after removal. The characteristic variables of the first training data after the second elimination are characteristic variables 1-3, 7-21, and 26-27. The elimination operation for other first training data is carried out in this way.

[0061] After obtaining each first training data after elimination, execute ④ in step (2). In ④, each second training data is determined according to each first training data after elimination, and a second training data set is obtained according to each second training data. Specifically, if one first training data after elimination is one second training data, each second training data is formed to obtain a second training data set.

[0062] In this embodiment, since the correlation coefficient between the designated characteristic variables in the SCADA data set (i.e., the first training data set) and most of the other characteristic variables is low, the designated characteristic variables are considered to have little effect on the training and prediction capabilities of the model. This is because they are relatively random relative to other characteristic variables, and the remaining characteristic variables are selected as input features of the model. Therefore, through feature selection processing, the designated characteristic variables in the first training data set are screened out, and then the designated characteristic variables of each first training data are eliminated to obtain each second training data, and then the second training data set is obtained.

[0063] The feature variables in the second training data set are more concentrated and have stronger predictive power, ensuring the comprehensiveness of the data set and facilitating the model training process and detection process. By focusing on the most informative feature variables, the complexity of the model is effectively simplified, potentially enhancing the performance of the model. The efficiency, detection precision and accuracy of the model training process and detection process are improved.

[0064] This embodiment takes into account the sequence length of the input data of the detection model. Determining the appropriate sequence length of input data based on the dimension of the time series is a key factor in the detection model's analysis of the time series data set. This balance is very important because input data with a sequence length that is too short may not capture enough time features, while input data with a sequence length that is too long may introduce unnecessary noise, exceeding the effective learning ability of the model. Therefore, step (3) is executed.

[0065] Step (3) performs time series division processing on the second training data set to obtain a sub-training set of wind turbines, wherein the time series division processing is a process of dividing the second training data set according to a preset time step. The preset time step can be set according to actual needs. Taking the second training data set shown in Table 6 as an example, the sampling interval is about 6s (seconds) and the sequence length is 100, which is equivalent to obtaining a sampling time length with a preset time step of about 10 minutes. According to the preset time step of about 10 minutes, the second training data numbered 1-100 are divided into a sub-training set, the second training data numbered 101-199 are divided into a sub-training set, and so on, to obtain a sub-training set of wind turbines.

[0066] Table 6

[0067]

[0068] During the model usage phase (i.e., the model inference phase), the sampling interval and sequence length can be set in advance, and then the preset time step can be set in advance. In this way, a data can be collected at a sampling interval (e.g., 6s), so that the test data of each wind turbine's test data set is more uniform in the time dimension. Then, according to the sequence length (e.g., the sequence length is set to 100), the test data set of each wind turbine is divided into sub-data sets with a preset time step of 10 minutes, so that the test data input into the model is more uniform and smooth. This enables the model to effectively learn from the data while capturing key time information, promoting accurate and meaningful time series analysis, further improving the detection precision and accuracy of wind turbines in a wide range of wind farms, and improving the recognition and detection efficiency of wind turbines in a wide range of wind farms.

[0069] In this embodiment, in order to effectively capture the time dynamic characteristics of blade failures of multiple wind turbines, the SCADA data is converted into a time series format, so that the detection model can accurately learn, identify and detect data based on time feature variables, improve the generalization ability and recognition detection efficiency of the model, and improve the detection precision and accuracy. Through time series partitioning processing, the second training data set is divided to obtain sub-training sets, which provides a sequence that is neither too short nor too long for model training, ensuring the best balance, enabling the model to effectively learn from the data, while capturing key time information, and improving the generalization ability and detection ability of the model.

[0070] The training set of each wind turbine executes the above steps (1)-(3) to obtain a sub-training set of each wind turbine, and then obtains multiple sub-training sets. After obtaining multiple sub-training sets, execute step 2 to normalize the multiple sub-training sets to obtain multiple sub-target training sets. The normalization process is to normalize each feature variable of the sub-training data in each sub-training set to obtain the target sub-training data. Then, each sub-training data in each sub-training data set executes the above operation of obtaining the target sub-training data to obtain the target sub-training set corresponding to each sub-training data set, and then obtain multiple target sub-training sets.

[0071] Specifically, for each sub-training data of each sub-training set, the characteristic variables of the sub-training data are normalized by the normalization technology of Z-score standardization to obtain the target sub-training data. Among them, the normalization technology can be set to other normalization technologies according to actual needs. Z-score standardization is shown in formula (2):

[0072] (2);

[0073] In the training phase, n2 represents the total number of feature variables of the sub-training data, t2 represents the index of the feature variable of the sub-training data, represents the target sub-training data, represents the characteristic variables of the sub-training data, Represents the mean value of the feature variable of the sub-training data. In addition, Z-score standardization is used to adjust each data point so that the mean of the feature variable of each data point is 0 and the standard deviation is 1.

[0074] In the inference phase, n2 represents the total number of feature variables of the sub-data, t2 represents the index of the feature variable of the sub-data, Indicates the target sub-data, represents the characteristic variable of the sub-data, Represents the average value of the characteristic variable of the sub-data.

[0075] The above operation of obtaining target sub-training data is performed on each sub-training data set to obtain a target sub-training set corresponding to each sub-training data set. Each target sub-training set is formed to obtain multiple target sub-training sets.

[0076] In this embodiment, it is crucial to normalize the input data of the model before starting the model training process to ensure that the characteristic variables of each data (i.e., the characteristic variables of each target sub-training set) are at a consistent scale so that the difference in measurement units will not affect the analysis and prediction results of the model. For example, if the units of temperature and wind speed are inconsistent, their measurement scales need to be unified to avoid the influence of model prediction due to the difference in measurement units.

[0077] After obtaining multiple target sub-training sets, execute step three, perform iterative training and detection processing on multiple sub-target training sets through the initial detection model, obtain the blade fault prediction result of each wind turbine, and converge the initial detection model through the focal loss function.

[0078] Specifically, Figure 2 As shown in FIG. 1 , the detection model includes: N bidirectional long short-term memory network layers BiLSTM layers, N-1 random dropout layers and a fully connected layer, where N>1. Two adjacent BiLSTM layers are connected through a dropout layer, and the output end of the last BiLSTM layer among the N BiLSTM layers is connected to the fully connected layer.

[0079] The iterative training detection processing process is that for each target sub-training set, A1, the target sub-training set is processed through the current BiLSTM layer and the dropout layer to obtain the current target hidden state, and then the current target hidden state is processed through the next BiLSTM layer and the dropout layer to obtain the next target hidden state, until the last BiLSTM layer is passed to obtain the output target hidden state. A2, the output target hidden state is processed through the fully connected layer to obtain the prediction result vector. A3, according to the prediction result vector, the prediction result probability of the target sub-training set is obtained, and then the blade fault type information of the wind turbine corresponding to the target sub-data set is obtained. A4, after each target sub-data set performs the operation of obtaining the prediction result probability, that is, after each target sub-data set performs steps A1-A3, the blade fault type information of each wind turbine is obtained, and the detection model is converged through the focal loss function.

[0080] In order to extract features related to blade failure from preprocessed and normalized time series data (i.e., target sub-training set), it is also necessary to analyze the time correlation within these time series data. To this end, the LSTM network is used because it is good at capturing time features. In order to enhance the capabilities of LSTM, this embodiment incorporates a bidirectional LSTM (i.e., BiLSTM layer) into the model so that it can analyze forward and backward time series data simultaneously. This ensures that the feature variables extracted at any given moment are affected by past and future contexts, thereby providing a more comprehensive understanding of time dependencies. However, the complexity of the multi-layer BiLSTM structure increases the risk of overfitting and may affect the performance of the model on unseen data. To solve this problem, the dropout technology is combined with the multi-layer BiLSTM to randomly shut down neurons during the training process, effectively reducing overfitting. The end of the model is a fully connected layer that projects the hidden state to the dimension of the classification. In this way, the model architecture of this embodiment is a multi-layer bidirectional long short-term memory network, i.e., a multi-layer BiLSTM layer, which enhances a random drop strategy. The random drop strategy is a strategy of randomly omitting neurons during training. Through the design of this model, the model can avoid overfitting during the detection process and improve the generalization ability of fault diagnosis of wind turbines at the wind farm level. In addition, the model can be used to realize generalized state monitoring and detection of wind turbines under a wide range of wind farm settings, quickly locate blade faults of wind turbines, improve the effectiveness of the detection process and the efficiency of identification and detection, and enhance detection precision and accuracy.

[0081] Take an example to illustrate the specific process of steps A1 and A2. Assume that the model includes: 4 bidirectional long short-term memory network layers BiLSTM layers, 3 random dropout layers and a fully connected layer. Two adjacent BiLSTM layers are connected by a dropout layer, and the output of the last BiLSTM layer among the N BiLSTM layers is connected to the fully connected layer. The specific connection relationship of this model is that the output of the first BiLSTM layer is connected to the input of the second BiLSTM layer through the first dropout layer. The output of the second BiLSTM layer is connected to the input of the third BiLSTM layer through the second dropout layer. The output of the third BiLSTM layer is connected to the input of the fourth BiLSTM layer through the third dropout layer. The output of the fourth BiLSTM layer is connected to the input of the fifth BiLSTM layer through the fourth dropout layer. The output of the fifth BiLSTM layer is connected to the fully connected layer, and the fully connected layer outputs the result.

[0082] For each target sub-training set, the target sub-training set is processed by the first BiLSTM layer and the first dropout layer in turn to obtain the target hidden state output by the first dropout layer. At this time, the target hidden state output by the first dropout layer is the current target hidden state, and the current BiLSTM layer and dropout layer are the first BiLSTM layer and the first dropout layer. The next BiLSTM layer and dropout layer are the second BiLSTM layer and the second dropout layer.

[0083] The target hidden state output by the first dropout layer is processed by the second BiLSTM layer and the second dropout layer in turn to obtain the target hidden state output by the second dropout layer. At this point, the target hidden state output by the second dropout layer is the current target hidden state, and the current BiLSTM layer and dropout layer are the second BiLSTM layer and the second dropout layer. The next BiLSTM layer and dropout layer are the third BiLSTM layer and the third dropout layer. And so on for the subsequent process. Until the target hidden state output by the fourth dropout layer is processed by the fifth BiLSTM layer, the target hidden state output by the fifth BiLSTM layer is obtained, that is, the output target hidden state. The output target hidden state is processed by the fully connected layer to obtain the prediction result vector.

[0084] The layer processing process of each BiLSTM layer is consistent, and the layer processing process of each dropout layer is consistent. The process of passing the target sub-training set through the current BiLSTM layer and the dropout layer to obtain the current target hidden state (i.e., the layer processing process of passing through a single-layer BiLSTM layer and a single-layer dropout layer in sequence) is to first process each target feature variable in each target sub-training data in the target sub-training set through the LSTM unit in the current BiLSTM layer to obtain the current hidden state. Then, through the current dropout layer, the current hidden state and the Bernoulli random variable are multiplied to obtain the current target hidden state.

[0085] The specific process of processing each target feature variable in each target sub-training data in the target sub-training set through the LSTM unit in the current BiLSTM layer to obtain the current hidden state is as follows: for each target feature variable in each target sub-training data in the target sub-training set, B1, the target feature variable passes through an LSTM unit in the current BiLSTM layer to obtain the first hidden state. B2, the target feature variable passes through another LSTM unit in the current BiLSTM layer to obtain the second hidden state. B3, based on the first hidden state and the second hidden state, the hidden state of the target feature variable is obtained. B4, each target feature variable performs the above operation of obtaining the first hidden state and the second hidden state, that is, each target feature variable executes steps B1-B3 to obtain the hidden state of each target feature variable. B5, based on the hidden state of each target feature variable, the current hidden state is obtained.

[0086] Specifically, Figure 4 As shown, the target sub-training data includes n3 target feature variables, denoted as ( , , ..., , ..., ), represents the target feature variable, and t3 represents the index of the target feature variable. Each target feature variable of the target sub-training data executes steps B1-B3: the target feature variable The first hidden state is obtained by passing an LSTM unit in the BiLSTM layer . Target feature variable Passing through another LSTM unit in the BiLSTM layer, the second hidden state is obtained According to the target feature variable The first hidden state of and the second hidden state , get the target feature variable The hidden state of . Target feature variable The first hidden state is obtained by passing an LSTM unit in the BiLSTM layer . Target feature variable Passing through another LSTM unit in the BiLSTM layer, the second hidden state is obtained According to the target feature variable The first hidden state of and the second hidden state , get the target feature variable The hidden state of . And so on, the target feature variable The hidden state of .

[0087] Target feature variable The first hidden state is obtained by passing an LSTM unit in the BiLSTM layer . Target feature variable Passing through another LSTM unit in the BiLSTM layer, the second hidden state is obtained According to the target feature variable The first hidden state of and the second hidden state , get the target feature variable The hidden state of Until the last target feature vector is obtained The hidden state of . The target feature variable The first hidden state is obtained by passing an LSTM unit in the BiLSTM layer . Target feature variable Passing through another LSTM unit in the BiLSTM layer, the second hidden state is obtained According to the target feature variable The first hidden state of and the second hidden state , get the target feature variable The hidden state of .

[0088] After obtaining the hidden state of each target feature variable, the hidden states of these target feature variables form the hidden state of the target sub-training data, that is, .

[0089] The LSTM unit performs dual calculations on the time series data of the input multi-feature variables (i.e., the target sub-training data of the target sub-training set). Each LSTM unit calculates in the opposite direction and outputs a hidden state. In forward or backward order, adjacent LSTM units that process adjacent target feature variables are connected. For example, Figure 4 As shown, in forward order, the target feature variables are processed The output of the LSTM unit is connected to the input of the LSTM unit that processes the target feature variable, and so on. In the backward order, the target feature variable is processed The output of the LSTM unit is connected to process the target feature variable The input of the LSTM unit, and so on.

[0090] The processing process of each LSTM unit is: Figure 5 The processing target characteristic variables shown LSTM unit, represents the input gate, represents the forget gate, represents the cell gate, Represents the output gate. Regardless of the forward or backward order, the target feature variable corresponds to an LSTM unit that processes itself. , , and The subscript and The subscripts are consistent, indicating that the target feature variable is processed The corresponding gate in the LSTM unit. , , and The subscript t3 represents the processing target feature variable The index of the corresponding gate in the LSTM unit. Target feature variable The following calculations are performed sequentially in the LSTM unit:

[0091] (3);

[0092] (4);

[0093] (5);

[0094] (6);

[0095] (7);

[0096] ( ) (8);

[0097] in, represents the Hadamard product, is the Sigmoid function, and k represents a real number. For example, in formula (3), k= . and represents the weight matrix of the input gate, and Represents the bias matrix of the input gate. and represents the weight matrix of the forget gate, and Represents the bias matrix of the forget gate. and represents the weight matrix of the cell gate, and Represents the bias matrix for cell gates. and represents the weight matrix of the output gate, and represents the bias matrix of the output gate. Specifically, and are all 0. Through recursive calculation, the hidden state of the input time series data can be obtained. Figure 5 The * in denoted the corresponding part in formula (3)-(8), such as ,*express .

[0098] After each target sub-training data of the target sub-training set executes steps B1-B5, the hidden state of each target sub-training data of the target sub-training set is obtained, and then the current hidden state of the target sub-training set is obtained. . Indicates The BiLSTM layer of the layer, m represents the index of the target sub-training data.

[0099] After getting the current hidden state of the target sub-training set After that, the current hidden state and the Bernoulli random variable are multiplied through the current dropout layer to obtain the current target hidden state: the current hidden state of the target sub-training set It will go through a dropout process, during which Multiplied by , which means multiplying the current hidden state by the Bernoulli random variable to obtain the current target hidden state of the target sub-training set. represents a Bernoulli random variable with a probability of assigning some value in the current hidden state to 0 at a specified dropout rate.

[0100] In step A1, the target sub-training set is processed through the current BiLSTM layer and the dropout layer to obtain the current target hidden state, and then the current target hidden state is processed through the next BiLSTM layer and the dropout layer to obtain the next target hidden state, until it passes through the last BiLSTM layer to obtain the output target hidden state. Therefore, the target sub-training set is processed through each BiLSTM layer and the dropout layer combination layer in turn, until it passes through the last BiLSTM layer to obtain the output target hidden state The superscript f in H indicates the last BiLSTM layer, and r indicates the index of the hidden state H.

[0101] Then, execute step A2 to process the output target hidden state through the fully connected layer to obtain the prediction result vector. Specifically, the output target hidden state Input to the fully connected layer. The fully connected layer includes a rectified linear unit and a linear layer. The rectified linear unit is a ReLU function (Linear rectification function). The calculation process of the fully connected layer is:

[0102] ( ) (9);

[0103] Among them, A represents the weight matrix of the fully connected layer, T represents the transposed matrix of A, b is the bias vector, and Y is the prediction result vector. is the ReLU activation function.

[0104] The role of the fully connected layer is to convert the feature variables learned by the BiLSTM layer into the prediction probabilities required by specific applications.

[0105] Next, step A3 is performed to obtain the prediction result probability of the target sub-training set according to the prediction result vector, and then obtain the blade fault type information of the wind turbine corresponding to the target sub-data set.

[0106] The prediction result vector Y output by the fully connected layer is subjected to the softmax operation to generate the probability distribution P of the output blade fault type. Calculated by the following formula:

[0107] (10);

[0108] Among them, y q For each prediction result in Y, q is the index of the prediction result. for y q The corresponding probability, that is, the probability of the predicted result, each P is formed. It should also be noted that in the inference stage of the model, the prediction result vector Y output by the fully connected layer is also subjected to a softmax operation.

[0109] For example, assuming that the prediction result vector Y = (1, 2, 5, ..., 3), each element in Y is a prediction result, and a prediction result represents the prediction result corresponding to a blade fault type. If it is assumed that there are 20 blade fault types in this embodiment, there are 20 prediction results in the prediction result vector Y. After the prediction result vector Y is subjected to the softmax operation, that is, through formula (10), the probability distribution P of the output blade fault type is obtained, P = (0.1, 0.3, 0.25, ..., 0.4). The prediction result probability and probability distribution of the target sub-training set are obtained, indicating that the blade fault type information of the wind turbine corresponding to the target sub-training set at the sampling time of the target sub-training set is obtained.

[0110] In order to enhance the model's ability to recognize various states of wind turbine blades, continuous parameter updates through back propagation are essential. In this case, this embodiment uses a focal loss function instead of the traditional cross entropy (CE) loss. The focal loss function is particularly effective when dealing with difficult-to-classify samples. It assigns a smaller loss to simple, easy-to-classify samples, thereby solving the problem of class imbalance.

[0111] Next, step A4 is performed. After each target sub-dataset performs the operation of obtaining the prediction result probability, that is, after each target sub-dataset performs steps A1-A3, the detection model is converged by the focus loss function. The formula of the focus loss function is:

[0112] Loss( )=- log (11);

[0113] in, is the weight factor that adjusts the contribution of each blade fault type. is a focusing parameter used to reduce the relative loss of well-classified samples, thereby prioritizing challenging samples during learning. is the probability that the sample is the true label.

[0114] In this embodiment, the Adam optimizer is used in combination with the focal loss function to effectively update the model parameters. This embodiment uses the focal loss function instead of the standard cross entropy loss to solve the challenges of data concentration and extreme data imbalance, improve the performance of the model in a multi-classification environment, and ensure a more balanced learning result.

[0115] When the current loss function value corresponding to the iterative detection model is not greater than the loss function threshold, the current detection model is used as the target detection model. The loss function threshold can be set according to actual needs. Therefore, the target detection model is trained. The target detection model is obtained through extensive SCADA data training, and the structural design of the detection model enables the model to extract common features and enhance the feature extraction capability of the model. In this way, the blade fault detection of wind turbines at the wind farm level is realized through the model and method of this embodiment. Such a model will ensure improved reliability and applicability on various wind turbines, and improve the generalization ability of fault diagnosis of wind turbines at the wind farm level. In addition, it promotes the model to realize generalized state monitoring and detection of wind turbines under a wide range of wind farm settings, quickly locate blade faults of wind turbines, improve the effectiveness of the detection process and the efficiency of identification and detection, and enhance detection precision and accuracy.

[0116] In addition, in the inference stage of the target detection model, the prediction result probability and probability distribution of the target sub-dataset are obtained, and the probability distribution of the target sub-dataset is subjected to the argmax operation to obtain the blade fault type information of the wind turbine corresponding to the target sub-training set. The Argmax operation is used to select the index with the highest probability from the probability distribution, and the index with the highest probability corresponds to the blade fault type of the wind turbine at the sampling time of the target sub-dataset. For example, the probability distribution of a certain target sub-training set P = (0.1, 0.3, 0.25, 0.9, ..., 0.4), through the Argmax operation, it is determined that 0.9 is the largest in P, and 0.9 corresponds to a blade fault of the local crack type. It is then determined that at the sampling time of the target sub-training set, the blade fault type of the wind turbine corresponding to the target sub-training set is the local crack type.

[0117] Next, step S102 is executed to pre-process the data set to be tested of each wind turbine to obtain a plurality of sub-data sets.

[0118] Specifically, the preprocessing process is: in each wind turbine test data set, the wind turbine test data set is processed for missing values ​​to obtain a first data set, wherein the missing value processing is to remove the test data with missing values. The first data set is processed for feature selection to obtain a second data set. The second data set is processed for time series division to obtain sub-data sets of the wind turbine, wherein the time series division is a process of dividing the second data set according to a preset time step. The test data set of each wind turbine performs the above-mentioned operation of obtaining the sub-data sets of the wind turbine to obtain multiple sub-data sets.

[0119] The process and principle of the preprocessing in step S102 are consistent with the process and principle of preprocessing the training set in the training phase of the model (i.e., step 1). No further elaboration will be given here. The process and principle of missing value processing in the preprocessing are consistent with the process and principle of (1) in the above step 1. The missing value processing process is used to remove the test data instances with missing values ​​in the test data set.

[0120] The process and principle of feature selection in preprocessing are consistent with the process and principle of (2) in the above step 1. The specific process of performing feature selection on the first data set to obtain the second data set is to perform correlation processing on any two feature variables of each first data in the first data set to obtain the correlation matrix of each first data. Perform matrix averaging operation on the correlation matrix of each first data to obtain the correlation matrix of the first data set, and obtain the specified feature variable based on the correlation matrix of the first data set. Eliminate the specified feature variable from the feature variable of each first data to obtain each first data after elimination, wherein the feature variable of the first data after elimination is the feature variable of the first data after elimination. Determine each second data based on each first data after elimination, and obtain the second data set based on each second data.

[0121] The process and principle of time series division in preprocessing are consistent with the process and principle of (3) in the above step 1. The time series division process is a process of dividing the second data set according to a preset time step.

[0122] The preprocessing process proposed in this embodiment removes data with missing values ​​and useless information to ensure the integrity and comprehensiveness of the data and avoid data redundancy. Through missing value processing, feature selection processing and time series partitioning processing, the feature variables of each sub-data set are more concentrated and have stronger prediction capabilities. By focusing on the most informative feature variables, the detection process of the model is effectively simplified, the detection performance of the model is potentially enhanced, the efficiency, detection precision and accuracy of the model training process and detection process are improved, and the reliability and effectiveness of the method of this embodiment are enhanced.

[0123] Next, step S103 is executed to perform normalization processing on the multiple sub-datasets to obtain multiple target sub-datasets.

[0124] Specifically, the process and principle of the normalization processing in step S103 are consistent with the process and principle of normalizing the training set in the training phase of the model (i.e., step 2). No further elaboration is given here. The normalization processing process is: in each sub-data in each sub-data set, each characteristic variable of the sub-data is normalized to obtain the target sub-data. The above operation of obtaining the target sub-data is performed on each sub-data in each sub-data set to obtain the target sub-data set corresponding to each sub-data set, thereby obtaining multiple target sub-data sets.

[0125] In this embodiment, it is crucial to normalize the input data of the target detection model to ensure that the feature variables of each data (i.e., the feature variables of each target sub-dataset) are at a consistent scale so that the difference in measurement units does not affect the analysis and prediction results of the model.

[0126] Then, step S104 is executed to perform detection processing on the multiple target sub-datasets through the target detection model to obtain blade fault type information of each wind turbine.

[0127] Specifically, the process and principle of the detection processing in step S104 are consistent with the process and principle of the detection processing on the training set in the training stage of the model (ie, step 3), and will not be elaborated here.

[0128] The specific process of detection processing is: for each target sub-dataset, C1, the target sub-dataset is processed through the current BiLSTM layer and the dropout layer to obtain the current target hidden state, and then the current target hidden state is processed through the next BiLSTM layer and the dropout layer to obtain the next target hidden state, until the last BiLSTM layer is passed to obtain the output target hidden state. C2, the output target hidden state is processed through the fully connected layer to obtain the prediction result vector. C3, according to the prediction result vector, the prediction result probability of the target sub-dataset is obtained, and then the blade fault type information of the wind turbine corresponding to the target sub-dataset is obtained. C4, after each target sub-dataset performs the operation of obtaining the prediction result probability, that is, each target sub-dataset executes C1-C3 to obtain the blade fault type information of each wind turbine.

[0129] The process and principle of step C1 are consistent with the process and principle of step A1 in the iterative training detection process of the model in the training phase. The process and principle of step C2 are consistent with the process and principle of step A2 in the iterative training detection process of the model in the training phase. The process and principle of step C3 are consistent with the process and principle of step A3 in the iterative training detection process of the model in the training phase. The process and principle of step C4 are consistent with the process and principle of step A4 in the iterative training detection process of the model in the training phase. No further elaboration will be given here.

[0130] Step C1, the target sub-dataset is processed through the current BiLSTM layer and the dropout layer to obtain the current target hidden state. The specific process is: through the LSTM unit in the current BiLSTM layer, each target feature variable in each target sub-data in the target sub-dataset is processed to obtain the current hidden state. Through the current dropout layer, the current hidden state and the Bernoulli random variable are multiplied to obtain the current target hidden state.

[0131] Through the LSTM unit in the current BiLSTM layer, each target feature variable in each target sub-data in the target sub-dataset is processed to obtain the current hidden state. The specific process is: for each target feature variable in each target sub-data in the target sub-dataset, D1, the target feature variable passes through an LSTM unit in the current BiLSTM layer to obtain the first hidden state. D2, the target feature variable passes through another LSTM unit in the current BiLSTM layer to obtain the second hidden state. D3, according to the first hidden state and the second hidden state, the hidden state of the target feature variable is obtained. D4, each target feature variable performs the above operation of obtaining the first hidden state and the second hidden state, that is, each target feature variable executes steps D1-D3 to obtain the hidden state of each target feature variable. D5, according to the hidden state of each target feature variable, the current hidden state is obtained.

[0132] The process and principle of step D1 are consistent with the process and principle of step B1 in the iterative training detection processing of the model in the training stage. The process and principle of step D2 are consistent with the process and principle of step B2 in the iterative training detection processing of the model in the training stage. The process and principle of step D3 are consistent with the process and principle of step B3 in the iterative training detection processing of the model in the training stage. The process and principle of step D4 are consistent with the process and principle of step B4 in the iterative training detection processing of the model in the training stage. The process and principle of step D5 are consistent with the process and principle of step B5 in the iterative training detection processing of the model in the training stage. No further elaboration will be given here.

[0133] It should also be noted that the blade fault type information of each wind turbine includes sampling time, the number of the wind turbine, and the blade fault type of the wind turbine.

[0134] In order to extract features related to blade failure from preprocessed and normalized time series data (i.e., target sub-datasets), it is also necessary to analyze the time correlation within these time series data. To this end, an LSTM network is used because it is good at capturing time features. In order to enhance the capabilities of LSTM, this embodiment incorporates a bidirectional LSTM (i.e., BiLSTM layer) into the model so that it can simultaneously analyze forward and backward time series data. This ensures that the feature variables extracted at any given moment are affected by past and future contexts, thereby providing a more comprehensive understanding of time dependencies. However, the complexity of the multi-layer BiLSTM structure increases the risk of overfitting and may affect the performance of the model for unseen data. To solve this problem, the dropout technique is combined with the multi-layer BiLSTM to randomly shut down neurons during the training process, effectively reducing overfitting. The end of the model is a fully connected layer that projects the hidden state to the classification dimension. In this way, through the detection method of this model and this embodiment, overfitting is avoided during the detection process, and the generalization ability of fault diagnosis of wind turbines at the wind farm level is improved. Moreover, it promotes the realization of generalized condition monitoring and detection capabilities for wind turbines in a wide range of wind farm settings, quickly locates blade faults of wind turbines, improves the effectiveness of the detection process and the efficiency of identification detection, enhances the detection precision and accuracy, and improves the effectiveness and reliability of the detection process.

[0135] Embodiment 2

[0136] Based on the same inventive concept, the second embodiment of the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of any one of the above-mentioned wind farm-level wind turbine blade fault detection methods are implemented.

[0137] Embodiment 3

[0138] Based on the same inventive concept, the third embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the program implements the steps of any method of the wind farm-level wind turbine blade fault detection method described in the first embodiment above.

[0139] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0140] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A wind turbine blade fault detection method at a wind farm level, characterized in that: include: Obtain a target detection model of a wind turbine at a wind farm level and a data set to be tested for each wind turbine in a plurality of wind turbines, wherein the target detection model comprises: N bidirectional long short-term memory network layers BiLSTM layers, N-1 random dropout layers dropout layers and a fully connected layer, N>1, two adjacent BiLSTM layers are connected by a dropout layer, and the output end of the last BiLSTM layer in the N BiLSTM layers is connected to the fully connected layer, and the data set to be tested for each wind turbine is the SCADA data to be tested for each wind turbine; Preprocessing the data set to be tested of each wind turbine to obtain multiple sub-data sets, including: In the data set to be tested of each wind turbine, missing value processing is performed on the data set to be tested of the wind turbine to obtain a first data set, wherein the missing value processing is to remove the data to be tested with missing values, and the missing value processing is used to make the complete data to be tested retain the time interval characteristics after the data to be tested with missing values ​​are removed; wherein each first data in the first data set carries a sampling time to retain key time dynamic information, which provides a basis for calculating the interval between consecutive data points; Performing feature selection processing on the first data set to obtain a second data set; Performing time series division processing on the second data set to obtain a sub-data set of the wind turbine, wherein the time series division processing is a process of dividing the second data set according to a preset time step; The data set to be tested of each wind turbine performs the above operation of obtaining the sub-data set of the wind turbine to obtain the multiple sub-data sets; Normalizing the multiple sub-datasets to obtain multiple target sub-datasets, wherein the normalization process is used to ensure that the feature variables of each target sub-dataset are at a consistent scale so that the difference in measurement units does not affect the analysis and prediction results of the target detection model; Performing detection processing on the multiple target sub-datasets by using the target detection model to obtain blade fault type information of each wind turbine; The step of performing feature selection on the first data set to obtain a second data set includes: Perform correlation processing on any two characteristic variables of each first data in the first data set to obtain a correlation matrix of each first data; Performing a matrix averaging operation on the correlation matrix of each first data to obtain a correlation matrix of the first data set, and obtaining a designated feature variable according to the correlation matrix of the first data set, wherein the designated feature variable is a feature variable that has a weak correlation with feature variables other than the own feature variable; Eliminating the specified characteristic variable from the characteristic variables of each first data to obtain each first data after elimination, wherein the characteristic variable of the first data after elimination is the characteristic variable of the first data after elimination; Each second data is determined according to each eliminated first data, and the second data set is obtained according to each second data.

2. The method according to claim 1, characterized in that The detecting and processing the plurality of target sub-datasets by the target detection model to obtain blade fault type information of each wind turbine includes: For each target sub-dataset, the target sub-dataset is processed through the current BiLSTM layer and the dropout layer to obtain a current target hidden state, and then the current target hidden state is processed through the next BiLSTM layer and the dropout layer to obtain a next target hidden state, until the last BiLSTM layer is passed to obtain an output target hidden state; Processing the output target hidden state through the fully connected layer to obtain a prediction result vector; According to the prediction result vector, obtaining the prediction result probability of the target sub-data set, and then obtaining the blade fault type information of the wind turbine corresponding to the target sub-data set; After the operation of obtaining the prediction result probability is performed on each target sub-data set, the blade fault type information of each wind turbine is obtained.

3. The method according to claim 2, characterized in that The step of processing the target sub-dataset through the current BiLSTM layer and the dropout layer to obtain the current target hidden state includes: Processing each target feature variable in each target sub-data in the target sub-data set through the LSTM unit in the current BiLSTM layer to obtain a current hidden state; The current target hidden state is obtained by multiplying the current hidden state and the Bernoulli random variable through the current dropout layer.

4. The method according to claim 3, characterized in that The LSTM unit in the current BiLSTM layer processes each target feature variable in each target sub-data in the target sub-data set to obtain a current hidden state, including: For each target feature variable in each target sub-data in the target sub-dataset, the target feature variable obtains a first hidden state through an LSTM unit in the current BiLSTM layer; The target feature variable passes through another LSTM unit in the current BiLSTM layer to obtain a second hidden state; Obtaining a hidden state of the target feature variable according to the first hidden state and the second hidden state; Each of the target feature variables performs the above operations of obtaining the first hidden state and the second hidden state to obtain the hidden state of each of the target feature variables; The current hidden state is obtained according to the hidden state of each of the target feature variables.

5. The method according to claim 1, characterized in that The normalizing the multiple sub-datasets to obtain multiple target sub-datasets includes: In each sub-data in each of the sub-data sets, each characteristic variable of the sub-data is normalized to obtain target sub-data; The above operation of obtaining the target sub-data is performed on each sub-data in each of the sub-data sets to obtain the target sub-data set corresponding to each of the sub-data sets, thereby obtaining the multiple target sub-data sets.

6. The method according to claim 1, characterized in that Acquiring the target detection model includes: The initial detection model is iteratively trained using a training set of each wind turbine among the multiple wind turbines that have been acquired, until a current loss function value corresponding to the iterative detection model is no greater than a loss function threshold, and the current detection model is used as the target detection model, wherein the current loss function value of the iterative detection model is obtained through a focal loss function.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the wind farm-level wind turbine blade fault detection method according to any one of claims 1 to 6 are implemented.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the wind farm-level wind turbine blade fault detection method according to any one of claims 1 to 6 are implemented.

Images