Wind turbine fault detection method based on time series confrontation

Abstract

The present application belongs to the field of wind turbine fault detection, and particularly relates to a wind turbine fault detection method based on time sequence confrontation, comprising: obtaining wind turbine data, pre-processing the wind turbine data, which comprises data dimension reduction and MinMax standardization processing on the wind turbine data; inputting the pre-processed wind turbine data into a trained wind turbine fault detection model based on time sequence confrontation; obtaining n base models through n training data respectively; then obtaining an abnormal score matrix through an integrated discriminant framework; and then obtaining a fault detection result through the abnormal score matrix; wherein the wind turbine fault detection model based on time sequence confrontation is composed of an encoder, a first decoder and a second decoder; in the process of conducting confrontation training on the first decoder and the second decoder, corresponding regularization terms are introduced into the objective functions of the generator and the discriminator to improve the gradient descent speed, thereby reducing the model training time.

Description

Technical Field

[0001] This invention belongs to the field of wind turbine fault detection, specifically relating to a wind turbine fault detection method based on time-series countermeasures. Background Technology

[0002] With the increasing global energy demand and the severe environmental pollution caused by fossil fuels, wind energy, as one of the most promising renewable energy sources, is receiving increasing attention from governments worldwide. According to the Global Wind Energy Council (GWEC), 8.8 GW of new offshore wind power capacity was added globally in 2022, accounting for 11.3% of the total new wind power capacity. Compared to onshore wind farms, offshore wind farms operate in unique environments with high salt spray and high humidity, making them more prone to failure. Early warning systems for faults can significantly reduce the operation and maintenance costs of offshore wind turbines. Therefore, it is necessary to conduct research on fault detection for offshore wind turbines, using data mining from existing SCADA systems to achieve accurate early fault detection, which is of great significance for improving the power generation efficiency of offshore wind farms.

[0003] Currently, wind turbine fault detection methods include statistical theory-based methods and machine learning-based methods. Statistical theory-based methods include cloud models and multinomial regression fitting; these methods offer good interpretability. However, due to the high dimensionality and strong correlation of parameters describing the operating status of wind turbines in SCADA systems, these methods face challenges such as difficult model construction and heavy computational burden. Machine learning-based methods primarily focus on the inherent characteristics and structure of the data, enabling the extraction of deeper information, preserving correlations between data points, and effectively handling high-dimensional variables. They are effective tools for early fault detection in wind turbines. For example, shallow neural networks are a typical example of machine learning, capable of quickly providing early warnings of wind turbine faults. However, their ability to extract feature information is limited, making it difficult to accurately identify early fault characteristics in wind turbines.

[0004] In recent years, deep neural networks have been widely applied to wind turbine fault detection. Since early-stage fault samples of wind turbines are extremely rare in practice, existing methods typically employ unsupervised learning, using deep neural networks to mine the reconstruction error between normal and abnormal data to determine wind turbine faults. In unsupervised mode, autoencoders and deep belief networks are commonly used deep neural networks. However, wind turbine operating data has a temporal sequence, making it difficult for the aforementioned models to extract its temporal features. Bi-LSTM extracts temporal features from both positive and negative directions by constructing two LSTMs. While Bi-LSTM can consider the temporal factors of the data, the model lacks robustness, is prone to overfitting, and still has limitations in handling long-term time dependencies. Summary of the Invention

[0005] To address the issues of inaccurate fault detection and untimely early warning in existing offshore wind turbine systems, inspired by generative adversarial networks (GANs), this invention proposes a wind turbine fault detection method based on temporal adversarial mechanisms. The fault detection model comprises a temporal adversarial base model for wind turbine fault detection and an integrated fault discrimination framework for wind turbines. The specific process includes: acquiring wind turbine data; preprocessing the wind turbine data, including dimensionality reduction and MinMax normalization; inputting the preprocessed wind turbine data into the trained temporal adversarial-based wind turbine fault detection model; obtaining n base models from n training data; then obtaining an anomaly score matrix through the integrated discrimination framework; and finally obtaining the fault detection result through the anomaly score matrix. The temporal adversarial-based wind turbine fault detection model consists of an encoder, a first decoder, and a second decoder.

[0006] The training process for a time-based adversarial wind turbine fault detection model includes:

[0007] S1: Obtain historical wind turbine data and preprocess the wind turbine data; the wind turbine data includes pitch motor temperature, generator speed, generator inlet water temperature, grid voltage and grid current;

[0008] S2: Perform data dimensionality reduction and MinMax standardization on wind turbine data;

[0009] S3: Input the preprocessed wind turbine data into the encoder for feature extraction, and compress the extracted features into a low-dimensional feature vector;

[0010] S4: Input the low-dimensional feature vectors into the first decoder and the second decoder respectively for data reconstruction, and perform adversarial training on the first decoder and the second decoder;

[0011] S5: Calculate the loss function of the model based on the reconstructed data, continuously adjust the model parameters, and complete the model training when the loss function converges.

[0012] S6: Integrate n trained models, input the wind turbine data to be tested into the n models, output the anomaly score matrix, and use the anomaly score matrix to determine the fault.

[0013] The preferred formula for dimensionality reduction of wind turbine data is:

[0014]

[0015]

[0016] Where σ is the standard deviation of each feature in the original data, N is the dimension of the original data, and Cov(x) i ,x j Find the covariance of X as a feature; calculate the covariance matrix of X; find the eigenvalues ​​and corresponding eigenvectors from the covariance matrix, and take the first M rows to form matrix P. Dimensionality reduction is performed on the data based on matrix P, using the following formula:

[0017] Y = PX

[0018] Where Y represents the data after being reduced from N dimensions to M dimensions, and X represents the original data.

[0019] The preferred formula for MinMax standardization of wind turbine data is:

[0020]

[0021] Where X′ represents the preprocessed wind turbine features, and x represents the original input wind turbine features.

[0022] Preferably, the encoder consists of one self-attention layer, three Bi-LSTM layers, and one fully connected layer; wherein the self-attention layer is used to weight the input wind turbine data; the data processed by the self-attention layer is input into the three Bi-LSTM layers to extract the temporal feature information; the temporal feature information is input into the fully connected layer to obtain a low-dimensional feature vector.

[0023] Furthermore, the formula for calculating the self-attention layer is as follows:

[0024]

[0025] x′ i =α i,j *x i

[0026] Where x i x j This represents the elements in the i-th and j-th columns of the input data, where element α i,j For x i and x j The similarity between them.

[0027] Furthermore, the calculation formula for the Bi-LSTM layer is as follows:

[0028] f t =σ(W f ·x′ t +b f )

[0029] i t =σ(W i ·x′ t +bi )

[0030]

[0031]

[0032] o t =σ(W o ·x′ t +b o )

[0033] h t =o t *tanh(C)

[0034] h = concat(h) L ,h R )

[0035] Among them, W f W i W c W o Both are learnable weight parameter matrices, b f b i b c b o All are corresponding biases, x t Let C be the wind turbine data at time t. i The wind turbine characteristic information at time t. This represents the temporary wind turbine characteristic information at time t.

[0036] Preferably, the structures of the first decoder and the second decoder are the same as those of the encoder.

[0037] The beneficial effects of this invention are:

[0038] Inspired by generative adversarial systems, this invention employs a two-stage training process for the first and second decoders in the model, thereby improving the model's robustness. The proposed encoder structure enables stable adversarial training, addressing the non-convergence problem commonly found in traditional generative adversarial networks. During adversarial training of the first and second decoders, this invention introduces corresponding regularization terms into the objective functions of the generator and discriminator to accelerate gradient descent and reduce model training time. Furthermore, this invention utilizes ensemble learning to comprehensively assess faulty wind turbines by integrating the anomaly score matrix output by the discriminant framework, avoiding the bias inherent in individual model judgments. Attached Figure Description

[0039] Figure 1 This is a structural diagram of the wind turbine fault detection model based on temporal adversarial mechanism of the present invention.

[0040] Figure 2 This is a flowchart of the fault detection based on ensemble learning according to the present invention. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] This invention proposes a temporal adversarial basic model for wind turbine fault detection. The model first obtains the temporal feature information and long-term dependencies of wind turbine data through feature mining technology based on Bi-LSTM and self-attention mechanism. Then, it uses a wind turbine fault detection model based on temporal adversarial mechanism to improve the robustness of the model. Subsequently, in order to avoid the one-sidedness of the output of a single model, this invention combines multiple basic models through an ensemble learning method, and further proposes a fault detection framework based on ensemble learning.

[0043] A wind turbine fault detection method based on temporal adversarial analysis is disclosed. The method includes: acquiring wind turbine data; preprocessing the wind turbine data, including dimensionality reduction and MinMax standardization; inputting the preprocessed wind turbine data into n trained temporal adversarial wind turbine fault detection models; and performing anomaly score detection by integrating a fault discrimination framework to obtain the fault detection result. The temporal adversarial wind turbine fault detection model consists of an encoder, a first decoder, and a second decoder.

[0044] The ensemble learning method is a bagging algorithm, which trains multiple base models. The fault determination method of the ensemble model is a voting method, which integrates the judgment results of multiple base models. If a wind turbine is judged as a faulty wind turbine more than half of the total number of base models, then the wind turbine is finally defined as a faulty wind turbine.

[0045] The training process for a time-based adversarial wind turbine fault detection model includes:

[0046] S1: Obtain historical wind turbine data and preprocess the wind turbine data; the wind turbine data includes pitch motor temperature, generator speed, generator inlet water temperature, grid voltage and grid current;

[0047] S2: Perform data dimensionality reduction and MinMax standardization on wind turbine data;

[0048] S3: Input the preprocessed wind turbine data into the encoder for feature extraction, and compress the extracted features into a low-dimensional feature vector;

[0049] S4: Input the low-dimensional feature vectors into the first decoder and the second decoder respectively for data reconstruction, and perform adversarial training on the first decoder and the second decoder;

[0050] S5: Calculate the loss function of the model based on the reconstructed data, continuously adjust the model parameters, and complete the model training when the loss function converges.

[0051] S6: Input the wind turbine data to be tested into n time-series adversarial wind turbine fault detection models after training, output the anomaly score matrix, and use the anomaly score matrix to determine the fault.

[0052] In one embodiment, the structure of the wind turbine fault detection model based on temporal adversarial mechanisms is as follows: Figure 1 As shown, the system includes one encoder and two decoders: AE1 is the generator, and AE2 is the discriminator. Wind turbine operating data can be described by variables such as pitch motor temperature, generator speed, generator inlet water temperature, grid voltage, and grid current. These variables can be acquired in real time through a Supervisory Control and Data Acquisition (SCADA) system. The initial collected data features number in the hundreds. The formula for dimensionality reduction of the wind turbine data is as follows:

[0053]

[0054]

[0055] Where σ is the standard deviation of each feature in the original data, N is the dimension of the original data, and Cov(x) i ,x j Given the covariance of features X, calculate the covariance matrix of X. Then, use the covariance matrix to find the eigenvalues ​​and corresponding eigenvectors, and take the first M rows to form matrix P.

[0056] The data after dimensionality reduction is as follows:

[0057] Y = PX

[0058] Where Y represents the data after being reduced from N dimensions to M dimensions, and X represents the original data.

[0059] To standardize data dimensions, wind turbine data X needs to be standardized using MinMax. The formula for MinMax standardization is:

[0060]

[0061] Where x′ represents the preprocessed wind turbine features. x represents the original input wind turbine features, such as generator speed, generator inlet water temperature, grid voltage, and grid current, x∈R N×1 Where N is the feature dimension. After preprocessing, each feature of the wind turbine operation data is between [0,1]. X′ is the feature matrix composed of each wind turbine feature x′, X′∈R N×M M is the characteristic number.

[0062] The processed wind turbine data X′ is input into the encoder to extract features and compress them into a low-dimensional feature vector Z. The encoder consists of one self-attention layer, three Bi-LSTM layers, and one fully connected layer. The attention layer allows each wind turbine feature to learn weighted information from other wind turbine features, thus better capturing the impact of correlations between SCADA data on early faults. The calculation formula for the attention layer is as follows:

[0063]

[0064] MultiHead(X′,X′,X′)=Concat(head1,…,head n W X′

[0065]

[0066] Where, d X′ Let X′ be the dimension. This is a learnable weight parameter matrix, and Concat(*) is the matrix merging operation.

[0067] In this embodiment, the weights of the attention layer can also be obtained from the similarity between two data points, and the data are weighted according to the calculated similarity weights. The formula for calculating the similarity between two data points is as follows:

[0068]

[0069] x′ i =α i,j *x i

[0070] Where x i x j These represent the elements in the i-th and j-th columns of the input data, respectively, with element α. i,j For x i and x j The similarity between them.

[0071] After passing through the attention layer, the wind turbine data is input into a stacked Bi-LSTM layer to extract temporal feature information. The Bi-LSTM is composed of two LSTMs arranged in opposite directions. Each LSTM consists of three gating units: a forget gate, a forgetting gate, and a forgetting gate. t Memory Gate t Output gate o t composition.

[0072] The process of Bi-LSTM can be summarized as follows: forgetting the wind turbine feature information from the previous time step, then memorizing the wind turbine feature information input at the current time step, and passing the feature information from the current time step to the next time step. Finally, at each time step, the hidden state h is output. i Finally, the forward and backward LSTM outputs are concatenated using the `concat(*)` function. The calculation formula for Bi-LSTM is as follows:

[0073] f t =σ(W f ·x′ t +b f )

[0074] i t =σ(W i ·x′ t +b i )

[0075]

[0076]

[0077] o t =σ(W o ·x′ t +b o )

[0078] h t =o t *tanh(C t )

[0079] h = concat(h) L ,h R )

[0080] Among them, W f W i W c Wo 均 Let b be a learnable weight parameter matrix. f b i b c b o All are corresponding biases, x t Let C be the wind turbine data at time t.i The wind turbine characteristic information at time t. This represents the temporary wind turbine characteristic information at time t.

[0081] After passing through 3 layers of Bi-LSTM, a fully connected layer is then used to map the wind turbine data into a low-dimensional feature vector Z. The calculation formula for the fully connected layer is as follows:

[0082] Z = Relu(W) h ·h+B h )

[0083] Among them W h It is a learnable weight parameter matrix. B h That is the corresponding bias.

[0084] Preferably, the fully connected layer uses ReLU(*) as the activation function. The formula for calculating ReLU(*) is as follows:

[0085] Relu(x) = max(0, x)

[0086] The feature vector Z is then input into the first decoder and the second decoder for reconstruction. The goal of reconstruction is to restore Z to the input wind turbine data as accurately as possible. The network structure and calculation formulas of the first and second decoders are the same as those of the encoder. The training objectives of the reconstruction stage are as follows:

[0087]

[0088]

[0089] Where L(X′,AE(X′)) is the reconstruction error between the input data and the output data, T is the time length of the input data, and M is the number of features of the input data.

[0090] Inspired by generative adversarial thinking, after data reconstruction, the two autoencoders undergo adversarial training to improve the model's robustness. The wind turbine reconstruction data AE1(X′) generated by generator AE1 is then input into AE2. The reconstructed data generated by AE1 is then reconstructed by AE2. The goal of AE1 is to reduce the difference between the input X′ and the output of AE2, while the goal of AE2 is to increase the difference between X′ and the output of AE2. In other words, the training objective is:

[0091]

[0092] In summary, each autoencoder has two training objectives. AE1 first minimizes the reconstruction error of the wind turbine data X′, and then minimizes the difference between the reconstruction outputs of X′ and AE2. For AE2, AE2 first minimizes the reconstruction error of X′, and then maximizes the reconstruction error of the wind turbine data AE1 as much as possible. Combined, the loss functions for AE1 and AE2 are:

[0093]

[0094]

[0095] Here, n represents the number of training epochs. As n increases, the level of attention AE1 and AE2 pay to the two targets also changes. AE1(X′) represents the wind turbine data reconstructed by the first decoder, and AE2(X′) represents the wind turbine data reconstructed by the second decoder. It is worth noting that AE2 does not strictly act as a discriminator in the generative adversarial network. This is because if its input is the original data, AE2 will not increase the difference between the input data and the reconstructed data; on the contrary, it will reduce the reconstruction error between the two. However, if its input is the data reconstructed by AE1, then AE2 will increase the reconstruction error between the input data and the reconstructed data.

[0096] In the loss function, n represents the number of model training epochs. In anomaly detection, the anomaly score is defined as follows:

[0097] A(X′)=α‖X′-AE1(X′)‖2+β‖X′-AE2(AE1(X′))‖2

[0098] The parameters α + β = 1 are used to balance the degree of attention given to false positives and true positives in the wind turbine fault diagnosis results. If α is larger than β, it means we reduce our focus on false positives and true positives. In short, α < β indicates a high detection sensitivity model, and α > β indicates a low sensitivity model.

[0099] To avoid the bias of a single wind turbine fault detection model, this invention, based on the concept of ensemble learning and mimicking the Bagging ensemble learning structure, integrates multiple basic models and proposes an ensemble fault discrimination framework for wind turbines, such as... Figure 2 Then, the anomaly score matrix output by the integrated discrimination framework is used to determine whether the wind turbine has malfunctioned, including:

[0100] Step 1: Anomaly Score Matrix. Using n wind turbine data sets as n test data points, train n basic models using the above model training method. Input these n test data points into each basic model, and each basic model will output corresponding n anomaly scores, forming an n×1 reconstruction error vector. Concatenate the anomaly scores output by the n models to obtain an n×n anomaly score matrix.

[0101] Step 2: Sorting the Anomaly Score Matrix. Each column of the anomaly score matrix represents the anomaly score output by a base model for n test data points. The scores in each column are then sorted in ascending order, meaning the last few rows contain the larger anomaly scores. These anomaly scores are then replaced with the corresponding wind turbine numbers, resulting in an n×n index matrix. Each column represents n error indices sorted by score for a given base model.

[0102] Step 3: Wind Turbine Fault Diagnosis. Assume there are k fault datasets among the n test data sets. Count the number of times each wind turbine index appears in the index matrix from row nk to row n, index 'a'. If the number of times a wind turbine index appears exceeds half the total number of wind turbines n, then mark that wind turbine as faulty.

[0103]

[0104] In this embodiment, the present invention collected wind turbine operation data from a domestic offshore wind farm, involving a total of 8 wind turbine units. Each wind turbine's operation data contains 356 features, including voltage, current, pressure, and temperature of related components. The time span is two weeks, and the sampling frequency of the wind turbine data is 1 time per minute. To reduce the computational load of the model and improve the fault detection speed, 96 representative features were selected from the original features using a data dimensionality reduction method as input features for the fault detection model. Since the dataset features are of different types, the units and measurement standards of the data may differ. Therefore, it is usually necessary to perform preprocessing operations to unify the data dimensions, ensuring that different data have the same units and measurement standards. Thus, the present invention uses a data normalization method to unify the data dimensions, mapping the data to the range of 0 to 1. Among the collected data from the 8 wind turbine units, two wind turbine datasets contain fault data.

[0105] To verify the superiority of the method proposed in this invention, the model of this invention is compared and analyzed with USAD (Unsupervised Anomaly Detection on Multivariate Time Series) and bidirectional long short-term memory network models. The model composition of the method of this invention and the comparative methods (M1-M3) is shown below, where M3 is the method proposed in this invention.

[0106] M1: An ensemble model for wind turbine fault detection based on a bidirectional long short-term memory neural network. The base model uses a bidirectional long short-term memory neural network with four layers. The first layer has 96 neurons, the second layer has 6 neurons, the third layer has 12 neurons, and the fourth layer has 96 neurons. The initial learning rate is set to 0.001, and the Adam gradient descent algorithm is used as the optimizer. The base model is then integrated to form the ensemble model.

[0107] M2: An ensemble model for wind turbine fault detection based on USAD. The base model uses a multivariate time series unsupervised anomaly detection method (USAD). The USAD model uses a fully connected neural network with 5 layers: 96 neurons in the first layer, 48 in the second, 6 in the third, 48 in the fourth, and 96 in the fifth. The initial learning rate is 0.001, and the Adam gradient descent algorithm is used as the optimizer. The base model is then integrated to form the ensemble model.

[0108] M3: A temporal adversarial ensemble model for wind turbine fault detection. The model framework consists of one encoder and two decoders. Both the encoder and decoders utilize bidirectional long short-term memory neural networks. The model has seven hidden layers: the first layer is an attention mechanism layer; the second layer has 96 neurons; the third layer has 48 neurons; the fourth layer has 24 neurons; the fifth layer is also an attention mechanism layer; the sixth layer has 24 neurons; and the seventh layer has 96 neurons. The initial learning rate is set to 0.001, and the Adam gradient descent algorithm is used as the optimizer. The above basic models are then integrated to form an ensemble model.

[0109] In terms of evaluation metrics, classification accuracy (A), classification precision (P), classification recall (R), and F1 score are used to measure the model's accuracy in detecting wind turbine faults. Specifically, classification accuracy refers to the ratio of correctly predicted faulty wind turbines to the total number of wind turbines, and its calculation formula is as follows:

[0110]

[0111] Where A is the classification accuracy; TP is the number of correctly classified actual faulty fans; TN is the number of correctly classified normal fans; FN is the number of incorrectly classified actual faulty fans as normal fans; and FP is the number of incorrectly classified normal fans as faulty fans.

[0112] Classification accuracy refers to the ratio of correctly classified faulty wind turbines to the number of wind turbines identified as faulty by the model, i.e.:

[0113]

[0114] Classification recall rate refers to the ratio of correctly classified faulty fans to the actual number of faulty fans, i.e.:

[0115]

[0116] The F1 score is an evaluation metric for model classification performance; it is the harmonic mean of accuracy and recall. The F1 score ranges from 0 to 1, with a higher value indicating better classification performance. The formula is as follows:

[0117]

[0118] To illustrate the improvement in model accuracy by the proposed method, the following calculation examples are set: M1, M2, and M3 are trained using data from 8 wind turbines in an unsupervised manner, and the base model trained on each wind turbine uses the other 7 wind turbine datasets as the test set through cross-testing.

[0119] Table 1 shows that compared to M1 and M2, M3 improved accuracy, precision, and recall by 25%, 50%, and 50%, respectively. In a real wind turbine unit, the faulty turbines were turbine 1 and turbine 2. Both M1 and M2 showed false positives and false negatives. However, M3 correctly identified both turbines 1 and 2, without any false positives or false negatives, demonstrating the effectiveness of M3. Table 2 shows the detection results of the three methods for the two faulty turbines. It can be seen that M1 and M2 performed poorly in detecting faulty turbines.

[0120] Table 1. Performance comparison of ensemble models based on different learning strategies

[0121]

[0122] Table 2. Inspection results of faulty fans

[0123]

[0124]

[0125] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A wind turbine fault detection method based on time-series adversarial mechanisms, characterized in that, include: Acquire wind turbine data and preprocess the wind turbine data, including data dimensionality reduction and MinMax standardization. The preprocessed data of n wind turbine units are input into n trained time-series adversarial wind turbine fault detection models, and anomaly scores are obtained by integrating the fault discrimination framework to obtain fault detection results. The wind turbine fault detection model based on temporal adversarial analysis consists of an encoder, a first decoder, and a second decoder. The training process for a time-based adversarial wind turbine fault detection model includes: S1: Acquire historical wind turbine data and preprocess the wind turbine data; the wind turbine data includes pitch motor temperature, generator speed, generator inlet water temperature, grid voltage and grid current; S2: Perform data dimensionality reduction and MinMax standardization on wind turbine data; S3: Input the preprocessed wind turbine data into the encoder for feature extraction, and compress the extracted features into a low-dimensional feature vector; S4: Input the low-dimensional feature vectors into the first decoder and the second decoder respectively for data reconstruction, and perform adversarial training on the first decoder and the second decoder; The network structure and computation formulas of the first and second decoders are the same as those of the encoder; the training objectives of the reconstruction phase are as follows: ； ； in The reconstruction error between the input and output data is T, where T is the time length of the input data and M is the number of features in the input data. For pre-trained wind turbine data, AE1 ( ) represents the wind turbine data reconstructed by the first decoder, AE2( This indicates the wind turbine data reconstructed by the second decoder. This indicates taking the average value. This represents the m-th eigenvector at time t; After the data reconstruction is complete, the two autoencoders will undergo adversarial training; the generator AE1 will generate the wind turbine reconstruction data AE1( The data generated by AE1 will be re-input into AE2, and the reconstructed data generated by AE1 will be reconstructed by AE2 again; the goal of AE1 is to reduce the input... The difference between AE2 and AE2 output; while the goal of AE2 is to increase... The difference between the output of AE2 and the output of AE2, i.e., the training objective, is: ； Among them, AE1 ( ) represents the wind turbine data reconstructed by the first decoder, AE2(AE1( This indicates that the wind turbine data reconstructed by the first decoder is re-inputted into the model and then reconstructed by the encoder and the second decoder. S5: Calculate the loss function of the model based on the reconstructed data, continuously adjust the model parameters, and complete the model training when the loss function converges. S6: Input the wind turbine data to be tested into n time-series adversarial wind turbine fault detection models after training, output the anomaly score matrix, and use the anomaly score matrix to determine the fault.

2. The wind turbine fault detection method based on time-series adversarial analysis according to claim 1, characterized in that, The formula for MinMax standardization of wind turbine data is: ； in, 'x' represents the preprocessed fan features, and 'x' represents the original input fan features.

3. The wind turbine fault detection method based on time-series countermeasures according to claim 1, characterized in that, The encoder consists of one self-attention layer, three Bi-LSTM layers, and one fully connected layer; The self-attention layer is used to weight the input wind turbine data; the data processed by the self-attention layer is input into a 3-layer Bi-LSTM to extract the temporal feature information; the temporal feature information is input into a fully connected layer to obtain a low-dimensional feature vector.

4. The wind turbine fault detection method based on time-series countermeasures according to claim 3, characterized in that, The formula for calculating the self-attention layer is: ； ； ； Where, d X′ for Dimensions , , This is a learnable weight parameter matrix, and Concat(*) is the matrix merging operation.

5. The wind turbine fault detection method based on time-series countermeasures according to claim 3, characterized in that, Bi The calculation formula for the LSTM layer is: ； ； ； ； ； ； ； Among them, W f W i W c W o Both are learnable weight parameter matrices, b f b i b c b o All are corresponding biases, x t The data for the wind turbine at time t. The wind turbine characteristic information at time t. This represents the temporary wind turbine characteristic information at time t.

6. The wind turbine fault detection method based on time-series countermeasures according to claim 1, characterized in that, The process of adversarial training of the first decoder and the second decoder includes a reconstruction training phase and an adversarial training phase. During the reconstruction training phase, the first and second decoders reconstruct the low-dimensional feature vectors generated by the encoder. The training objectives for the first decoder and the second decoder are constructed respectively; during the adversarial training phase, the reconstruction result of the first decoder is input into the encoder again and mapped into a low-dimensional feature vector. The low-dimensional feature vector of the mapping is input into the second decoder for reconstruction; the training objectives of the first decoder and the second decoder are constructed. Training is completed when the training objectives of the first decoder and the second decoder are optimal or the number of training iterations reaches the upper limit. The training objective of the first decoder is to reduce the difference between the input data and the output of the second decoder, while the training objective of the second decoder is to increase the difference between the input data and the output of the second decoder.

7. The wind turbine fault detection method based on time-series countermeasures according to claim 1, characterized in that, The model's loss function is: ； ； in, Let the loss function of the first decoder be... Let the loss function of the second decoder be... This refers to the number of training epochs for the model. It is a Euclidean norm. For the characteristics of the pre-treated fan, Wind turbine data reconstructed by the first decoder; The wind turbine data is reconstructed by two decoders. The first half of the loss function of each decoder is the reconstruction loss, and the second half is the adversarial training loss.

Images