Wind driven generator fault detection method based on multivariate data and deep learning

Abstract

The invention discloses a wind driven generator fault detection method based on multivariate data and deep learning, and the method comprises the steps: carrying out the data feature fusion and reconstruction of a large amount of multivariate time series data obtained by a common SCADA system of a wind driven generator set through a variational auto-encoder model; establishing and selecting a state space deep learning model as an identification model; and carrying out supervised training on the model by using reconstructed data, and then carrying out fault detection of wind power generation. According to the method, in order to solve the problem that a deep learning model is difficult to learn few sample faults due to fault data imbalance, VAE is adopted for data reconstruction, fault samples are increased, the recognition accuracy of a subsequent model is effectively improved, and meanwhile, the subsequent model can carry out multi-fault recognition through data feature fusion. In addition, the selection state space deep learning model structure proposed in the invention can improve the accuracy and efficiency of fault detection of the wind generating set.

Description

Technical Field

[0001] The present invention belongs to the fields of deep learning, fault detection, wind power generation and data feature fusion, and specifically relates to a wind turbine fault detection method based on multivariate data and deep learning. Background Art

[0002] As a renewable and clean energy, wind energy has been widely used around the world. Wind power generation is a clean and pollution-free way of utilizing renewable energy. It has the characteristics of clean and environmental protection, renewability, short infrastructure cycle, flexible installed capacity, and low operation and maintenance costs. In recent years, with the strong support of national policies for the development of renewable energy, the installed capacity of wind power generation has also shown a steady growth trend. In 2023, the cumulative installed capacity of wind power generation in China was 441.34 million kilowatts, a year-on-year increase of 20.7%. The latest data from the National Energy Administration shows that from January to February 2024, the cumulative installed capacity of wind power generation in China was 450.35 million kilowatts, a year-on-year increase of 21.3%. In order to achieve the reliability and stability of wind turbine operation, a state monitoring and fault diagnosis system for wind turbines must be established to reduce the impact of faults. The monitoring and data acquisition system collects the operating data and working conditions of wind turbines, such as wind speed, temperature, speed, etc., which have been widely used in the fault diagnosis of wind turbines. There are many methods for wind turbine fault diagnosis, which are mainly divided into physics-based methods and data-driven methods. Cambron et al. (P.Cambron, A.Tahan, C.Masson, F.Pelletier, Bearing temperature monitoring of a wind turbine using physics-based model, J.Qual.Mainten.Eng.23(4)(2017)479e488.) proposed a dynamic model based on the difference between measured temperature and estimated temperature for monitoring the main bearing of a wind turbine. However, this requires specific wind turbine information and is not generalizable. With the rise of deep learning, data-driven methods have shown great vitality. Yang et al. (Yang L, Zhang Z. Windturbine gearbox failure detection based on SCADA data: A deep learning-based approach [J]. IEEE Transactions on Instrumentation and Measurement, 2020, 70: 1-11.) extracted vibration signal variables from the SCADA system as a data set and trained a deep learning model for binary classification to determine whether the gearbox is faulty. However, these methods currently do not make sufficient use of SCADA data, and the models cannot solve the problem of unbalanced data distribution. Summary of the invention

[0003] The purpose of the present invention is to propose a wind turbine fault detection method based on data fusion reconstruction and deep learning. A VAE model is used to fuse and reconstruct data features for a large amount of multivariate time series data obtained by a SCADA system commonly used by wind turbines. A state space deep learning model is selected as a recognition model to construct a deep learning model suitable for the fused data. After training, wind turbine fault detection can be performed.

[0004] The present invention is achieved by at least one of the following technical solutions.

[0005] A wind turbine fault detection method based on multivariate data and deep learning comprises the following steps:

[0006] S1. SCADA system collects data of many different variables;

[0007] S2, preprocessing the collected data;

[0008] S3. Based on the data fusion and feature fusion method of variational autoencoder (VAE), a variational autoencoder model is constructed. Through data fusion and feature fusion, a data set that can be used for deep learning network learning is obtained, and the recognition effect of subsequent models is improved;

[0009] S4. Build a deep learning model based on state space selection (SSM), use the data set obtained in step S3 for training, and use the trained deep learning model for fault detection.

[0010] Furthermore, the data collected by SCADA includes the following five categories:

[0011] 1) Wind power variables, including instantaneous wind speed, average wind speed, average wind speed, and wind angle;

[0012] 2) Temperature variables, including cabin temperature, gearbox cooling water, gearbox oil, generator stator, high-speed bearing, low-speed bearing, drive end bearing, free end bearing, cooling fan, tower top, and tower bottom temperature;

[0013] 3) Electrical variables, including phase voltage, phase current, phase frequency, and power factor;

[0014] 4) Operating condition variables, including cabin position, hydraulic system pressure, cabin vibration, blade speed, generator speed, blade angle;

[0015] 5) Statistical variables, including daily power generation, monthly power generation, and annual power generation.

[0016] Furthermore, the data preprocessing includes data cleaning, data alignment, and data normalization, which specifically include the following contents:

[0017] Data cleaning: The data in SCADA is collected at fixed time intervals, the default values ​​and abnormal values ​​are filtered out, and the motor operation status at the current timestamp is obtained based on the timestamp of the filtered data;

[0018] Data alignment and data normalization: Label the cleaned data according to the running status. Note that the first h data of the abnormal running status are marked as the same abnormality. Normalize the data of each variable and separate the data according to the label.

[0019] Furthermore, the variational autoencoder model comprises:

[0020] Encoder: Preprocessed data from different sources and types are input into the encoder, which consists of several layers of neural networks, processing different types of data and mapping them into latent space;

[0021] Hidden space: Map the input data to the latent space to generate latent variables. The latent variables will not be directly used as the output of the model, but as the input of the decoder.

[0022] Decoder: Decodes the latent variable representation into data in the target format. The decoder consists of a multi-layer neural network to ensure that the output dimension is correct.

[0023] Furthermore, the VAE model training includes the following:

[0024] Determine the optimization goal. The purpose of the VAE model is to obtain data from the dataset X through the encoder and decoder. Its optimization goal is to maximize its likelihood function logp θ (X);

[0025] According to the total probability formula:

[0026] logp θ (X) = ∫ z q φ (z|X)logp θ (X)dz (1)

[0027] where p θ (X) is the probability density function of sample X, z represents the latent variable, q φ (z|X) represents the approximate posterior distribution;

[0028] By Bayes' theorem:

[0029]

[0030] where p θ(X,z) represents the joint probability distribution density of sample X and latent variable z, q φ (z|X) is the approximate posterior distribution of the encoder, p θ (z|X) is the distribution of the decoder; after deformation, we get:

[0031]

[0032] Will Denoted as e(p θ ,q φ ),for According to the definition of KL divergence;

[0033] If P and Q are two distributions, then their D KL The (P,Q) divergence formula is as follows:

[0034]

[0035] Where P(x) is the probability density or probability mass of distribution P at sample x, and Q(x) is the probability density or probability mass of distribution Q at x;

[0036] So we can get:

[0037] logp θ (X)==l(p θ ,q Φ )+D KL (q φ ,p θ ) (5)

[0038] KL divergence D KL (q φ ,p θ )>0, the optimization goal is to maximize e(p θ ,q φ );

[0039] According to Bayes' theorem:

[0040]

[0041] After deformation, we can get:

[0042]

[0043] in is the expectation, p(z) is the probability density or probability mass of distribution P at sample z, from q φ (z|X) is sampled m times Zj to approximate We can get:

[0044]

[0045] Where n represents the number of samples, p θ (x I ∣z J ) is the probability at a certain sampling time; Formula (8) is used as the loss function term to train the VAE model.

[0046] Furthermore, in step S3, obtaining a data set available for deep learning network learning includes:

[0047] The data with different labels are input into the VAE model for training respectively, and the data is reconstructed into (batch, feature, length) format through the encoder and decoder as the input of the subsequent model, where batch represents the batch size, feature represents the number of features, and length represents the length of the generated data;

[0048] According to the specific fault type or detection target, each time series segment is labeled accordingly;

[0049] The labeled data is divided into training set, validation set and test set for subsequent model training.

[0050] Furthermore, constructing a deep learning model based on state space selection includes:

[0051] The classical state space equation under continuous time conditions is:

[0052] h′(t)=Ah(t)+Bx(t) (9)

[0053] y(t)=Ch(t) (10)

[0054] Where h′(t) is the state change of the system in continuous time t, h(t) is the state vector of the system, A is the state matrix, B is the input matrix, C is the output matrix, x(t) and y(t) are the input and output vectors respectively;

[0055] Discretization:

[0056]

[0057] y t =Ch t (12)

[0058] in Calculated from A and B under the continuous system;

[0059] Using convolution form:

[0060]

[0061] in, is the state vector in convolutional form; A, B, and C are all fixed parameters. A selection mechanism is added to make the B and C matrices change with the input. This is to increase the dimensions of the B and C matrices, incorporate the sequence length into B and C, and use the Linear layer for selection. This allows the B and C matrices to dynamically change each time the input is different.

[0062] Furthermore, the deep learning model based on state-space selection is a state-space algorithm that adds a selection mechanism to the main loop of the gated attention architecture. The state-space algorithm is regarded as a deep learning layer, and a convolutional layer is added before it to establish dependencies between different word vectors. A linear layer, a normalization layer, an activation layer, and a selection mechanism SSM block are added to the deep neural network.

[0063] Furthermore, the deep learning model based on state space selection adds residual connections to the selection mechanism SSM block and performs normalization before each input to make the model depth greater. The SSM blocks are stacked, and the number of stacking times is determined by the performance of the equipment performing the training. A normalization layer, a linear layer, and a softmax layer are added after the last block for fault classification and identification.

[0064] Furthermore, the trained deep learning model is deployed locally or in the cloud, including:

[0065] Select a set of local or cloud devices that can be used to deploy VAE models and selection mechanism SSM models, deploy the trained models to relevant devices, and perform real-time wind turbine fault detection.

[0066] Compared with the existing technology, the advantages and beneficial effects of the present invention are:

[0067] The present invention proposes a complete set of wind turbine fault detection methods. In view of the large number of default values, abnormal values, and the difficulty in manually selecting features in the actual SCADA system, the VAE model is used in combination with the Resnet residual network to not only complete the self-selection of data features, but also fuse and align the data features. This method not only greatly simplifies the production of data sets, but also completes the preliminary feature fusion and extraction of data, and improves the accuracy of the subsequent discriminant model. Since VAE is a generative model, it can also solve the data imbalance problem that the fault data is far less than the normal data. The choice of a deep learning model based on state space selection as a model for fault detection has the following two advantages in addition to its high recognition accuracy:

[0068] Different from CNN (convolutional neural network) and RNN (recurrent neural network), the state space model has both fast training and fast reasoning capabilities, that is, it can make full use of the efficient parallel processing capabilities of CNN while maintaining the flexibility of RNN in sequence data processing.

[0069] Different from the state-space model, the parameters of the state-space model with the added selection mechanism depend on the input. The model can dynamically adjust its behavior according to different inputs, and the generalization ability of the model is enhanced. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 This is a network structure diagram of data fusion in this embodiment and using the data to train a deep learning model based on state space selection;

[0071] Figure 2 Schematic diagram of the VAE model structure of this embodiment;

[0072] Figure 3 This is a schematic diagram of the state space model algorithm for the selection mechanism of this embodiment;

[0073] Figure 4 This is a schematic diagram of the SSM block structure of the selection mechanism in this embodiment;

[0074] Figure 5 Schematic diagram of the identification model for this implementation. DETAILED DESCRIPTION

[0075] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0076] This embodiment discloses a method for detecting wind turbine faults using SCADA system data, based on VAE to generate fused data and using a state space model, which can effectively improve the efficiency of wind turbine data processing and the accuracy of wind turbine fault detection. The overall process is shown in the attached figure. Figure 1 As shown, the method mainly includes the following two parts:

[0077] 1. Construction of VAE model: Construct encoder and decoder, use reparameterization method to solve the problem of non-differentiability in training process, obtain the specific expression of VAE loss function based on the assumed distribution, input data for training, and finally obtain a new data set;

[0078] 2. Construction and training of state-space deep learning model with selection mechanism: By adding the selection mechanism to the SSM model, a network block based on state-space selection is constructed. According to the performance of the training and inference equipment, these blocks are stacked to obtain a fault detection model; the model is trained using the data set processed by VAE, and finally the trained model is deployed. The deployed model can be used for real-time fault and diagnosis of wind turbines.

[0079] The present invention will be further described below in conjunction with the embodiments and the accompanying drawings.

[0080] This example provides a wind turbine fault detection method based on multivariate data and deep learning. Figure 1 As shown, the specific steps include:

[0081] S1. Construction of VAE model;

[0082] The structure diagram of VAE is shown in the attached Figure 2 , the specific contents include:

[0083] S1.1. Encoder: The encoder is composed of a convolutional neural network. In particular, in this example, a one-dimensional Resnet block is used to build the encoder, and the mean and log-variance of the latent variables are generated through a fully connected layer.

[0084] S1.2. Decoder: The decoder is composed of a convolutional neural network. In particular, in this example, a one-dimensional Resnet block is used to build the decoder. Unlike the encoder, the convolution in the Resnet block of the decoder is a transposed convolution. The length recovery of the time series is achieved by combining the convolution layer with the reverse convolution operation.

[0085] S1.3. Input and output: According to the data collected by the SCADA system, generally, SCADA will collect a variety of different variables.

[0086] As a preferred embodiment, the collected variables can be roughly divided into 5 categories:

[0087] (1) Wind power variables, including instantaneous wind speed, 30-second average wind speed, 60-second average wind speed, wind angle, etc.;

[0088] (2) Temperature variables, including cabin temperature, gearbox cooling water, gearbox oil, generator stator, high-speed bearing, low-speed bearing, drive end bearing, free end bearing, cooling fan, tower top, tower bottom temperature, etc.;

[0089] (3) Electrical variables, including phase voltage, phase current, phase frequency, power factor, etc.

[0090] (4) Operating condition variables, including cabin position, hydraulic system pressure, cabin vibration, blade speed, generator speed, blade angle, etc.;

[0091] (5) Statistical variables, including daily power generation, monthly power generation, annual power generation, etc.

[0092] Assume that the number of variables that can be used as input is D, the length of the input time series is L, and there are B segments of such a time series. Then the format of the input data is (B, D, L), which means that there are B batches of data with a time series length of L, and each time series contains D variables.

[0093] The format of the output data is (B, F1, L1), where F1 is the desired number of features and L1 is the length of the output sequence. In this example, the output is (32, 5, 1000).

[0094] S1.4.VAE training: The purpose of the VAE model is to transform the dataset X through the encoder and decoder to obtain The data in is generated, but its distribution is close enough to X. Therefore, its optimization goal is to maximize its likelihood function logp θ (X).

[0095] According to the total probability formula:

[0096] logp θ (X) = ∫ z q φ (z|X)logp θ (X)dz (1)

[0097] By Bayes' theorem:

[0098]

[0099] where p θ (X,z) represents the joint probability distribution density of sample X and latent variable z, q φ (z|X) is the approximate posterior distribution of the encoder, p θ (z|X) is the decoder distribution.

[0100] After deformation, we can get:

[0101]

[0102] Let the first term be l(p θ ,q φ ), for the latter item, according to the definition of KL divergence:

[0103] KL divergence loss measures the difference between the latent variable distribution and the standard normal distribution. KL divergence is a method to measure the similarity between two probability distributions P and Q. It is asymmetric. If P and Q are two distributions, the KL divergence formula is as follows:

[0104]

[0105] Where P(x) is the probability density or probability mass of distribution P at sample x, and Q(x) is the probability density or probability mass of distribution Q at x.

[0106] So we can get

[0107] logp θ (X)==l(p θ ,q φ )+D KL (q φ ,p θ ) (5)

[0108] Since the KL divergence D KL (q φ ,p θ )>0, so the optimization goal is to maximize e(p θ ,q φ ).

[0109] According to Bayes' theorem

[0110]

[0111] After deformation, we can get:

[0112]

[0113] in For expectation, generally from q φ We sample Zj (the jth latent variable) m times in (z|X) to approximate We can get:

[0114]

[0115] Where n represents the number of samples, p θ (x i ∣z j ) is the probability at a specific sampling time.

[0116] In particular, we can assume that the distribution q φ The distribution p follows a Gaussian distribution. In order to facilitate training, we minimize -l(p θ ,q φ ), and let m = 1, that is, for each x i , only one z is sampledj , we can get:

[0117]

[0118] in μ i is the variance and mean of the latent variable z, corresponding to the parameter of the i-th sample, d is the dimension of the latent variable, μ i ′ is the generated reconstructed sample.

[0119] According to this formula, training optimization can be performed under the pytorch framework.

[0120] As a preferred embodiment, the data with different labels are re-input into the trained VAE, and the data is reconstructed into data in the (32, 5, 1000) format through the encoder and decoder as the input of the subsequent model;

[0121] According to the specific fault type or detection target, each time series segment is labeled accordingly.

[0122] The processed data is divided into training set, validation set and test set to ensure the fairness and effectiveness of model training and evaluation.

[0123] S2. Construct a deep learning model based on state space selection, as follows:

[0124] S2.1, add the selection mechanism into the SSM model;

[0125] The working principle of SSM is as follows:

[0126] The classical state space equation under continuous time conditions is:

[0127] h′(t)=Ah(t)+Bx(t) (8)

[0128] y(t)=Ch(t) (9)

[0129] Where h′(t) is the state change of the continuous-time system, h(t) is the state vector of the system, A is the state matrix, B is the input matrix, C is the output matrix, and x(t) and y(t) are the input and output vectors respectively.

[0130] Discretization:

[0131]

[0132] y t =Ch t (11)

[0133] in Calculated from A and B under the continuous system;

[0134] Using convolution form:

[0135]

[0136] in is the state vector in convolution form. The selection mechanism is added to SSM so that the A, B, and C matrices change with the input. Figure 3 As shown;

[0137] S2.2, construct the SSM block based on the selection mechanism;

[0138] In this example, in order to construct an SSM block based on the selection mechanism, such as Figure 4 As shown in the figure, the above SSM algorithm is added to the main loop of the gated attention architecture. This module contains a linear layer, a convolutional layer, a selective SSM, and an activation function Silu. Adding a convolutional layer before the SSM can establish dependencies between different tokens, preventing the subsequent calculation process from processing each token independently. This helps the model better capture the contextual information in the sequence, making the processing of sequence data more consistent and effective.

[0139] S2.3. Building a complete model

[0140] The complete SSM recognition model is composed of the stacked SSM blocks constructed in step S2.2, as shown in Figure 5 As shown. In order to use multiple blocks, residual connections are added inside each block. Finally, the Norm layer, Linear layer, and softmax layer are added at the end to obtain the output probability for completing the classification task, as shown in Figure 5 As shown. On the RTX4070 GPU, 4 blocks are recommended.

[0141] S2.4. Model training and reasoning

[0142] The data set (training set and validation set) obtained in step S1 is input into the model for training. The training framework uses pytorch and the loss uses the cross entropy loss function. The training is performed until the loss no longer decreases. The inference is performed using the test set. Note that it is recommended to use cosine annealing to adjust the learning rate and it is recommended to use learning rate warm-up. The final model can be used for fault detection of wind turbines.

[0143] S2.5. Model deployment and real-time fault detection

[0144] The models obtained in step S1 and step S2 are deployed in series on a local computing device or a cloud computing device, and the multivariate data collected by the system are input in real time to perform real-time wind turbine fault detection.

[0145] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. A wind turbine fault detection method based on multivariate data and deep learning, characterized in that: The following steps are involved: S1. SCADA system collects data of many different variables; S2, preprocessing the collected data; S3. Based on the data fusion and feature fusion method of variational autoencoder, a variational autoencoder model is constructed. Through data fusion and feature fusion, a data set that can be used for deep learning network learning is obtained, and the recognition effect of subsequent models is improved; S4. Build a deep learning model based on state space selection, use the data set obtained in step S3 for training, and use the trained deep learning model for fault detection.

2. A wind turbine fault detection method based on multivariate data and deep learning according to claim 1, characterized in that: The data collected by SCADA includes the following five categories: 1) Wind power variables, including instantaneous wind speed, average wind speed, average wind speed, and wind angle; 2) Temperature variables, including cabin temperature, gearbox cooling water, gearbox oil, generator stator, high-speed bearing, low-speed bearing, drive end bearing, free end bearing, cooling fan, tower top, and tower bottom temperature; 3) Electrical variables, including phase voltage, phase current, phase frequency, and power factor; 4) Operating condition variables, including cabin position, hydraulic system pressure, cabin vibration, blade speed, generator speed, blade angle; 5) Statistical variables, including daily power generation, monthly power generation, and annual power generation.

3. A wind turbine fault detection method based on multivariate data and deep learning according to claim 1, characterized in that: The data preprocessing includes data cleaning, data alignment, and data normalization, which specifically include the following contents: Data cleaning: The data in SCADA is collected at fixed time intervals, the default values ​​and abnormal values ​​are filtered out, and the motor operation status at the current timestamp is obtained based on the timestamp of the filtered data; Data alignment and data normalization: Label the cleaned data according to the running status. Note that the first h data of the abnormal running status are marked as the same abnormality. Normalize the data of each variable and separate the data according to the label.

4. A wind turbine fault detection method based on multivariate data and deep learning according to claim 1, characterized in that: The variational autoencoder model comprises: Encoder: Preprocessed data from different sources and types are input into the encoder, which consists of several layers of neural networks, processing different types of data and mapping them into latent space; Hidden space: Map the input data to the latent space to generate latent variables. Decoder: Decodes the latent variable representation into data in the target format. The decoder consists of a multi-layer neural network to ensure that the output dimension is correct.

5. A wind turbine fault detection method based on multivariate data and deep learning according to claim 4, characterized in that: Training VAE model training includes the following: Determine the optimization goal. The purpose of the VAE model is to obtain data from the dataset X through the encoder and decoder. Its optimization goal is to maximize its likelihood function logp θ (X); According to the total probability formula: logp θ (X)=∫ z q φ (z∣X)logp θ (X)dz (1) where p θ (X) is the probability density function of sample X, z represents the latent variable, q φ (z|X) represents the approximate posterior distribution; By Bayes' theorem: where p θ (X,z) represents the joint probability distribution density of sample X and latent variable z, q φ (z|X) is the approximate posterior distribution of the encoder, p θ (z|X) is the distribution of the decoder; after deformation, we get: Will Denoted as l(p θ ,q φ ),for According to the definition of KL divergence; If P and Q are two distributions, then their D KL The (P,Q) divergence formula is as follows: Where P(x) is the probability density or probability mass of distribution P at sample x, and Q(x) is the probability density or probability mass of distribution Q at x; So we can get: logp θ (X)==l(p θ ,q φ )+D KL (q Φ ,p θ ) (5) KL divergence D KL (q Φ ,p θ )>0, the optimization goal is to maximize l(p θ ,q φ ); According to Bayes' theorem: After deformation, we can get: in is the expectation, p(z) is the probability density or probability mass of distribution P at sample z, from q φ (z|X) is sampled m times Zj to approximate We can get: Where n represents the number of samples, p θ (x i ∣z j ) is the probability at a certain sampling time; Formula (8) is used as the loss function term to train the VAE model.

6. A wind turbine fault detection method based on multivariate data and deep learning according to claim 1, characterized in that: In step S3, obtaining a data set available for deep learning network learning includes: The data with different labels are input into the VAE model for training respectively, and the data is reconstructed into (batch, feature, length) format through the encoder and decoder as the input of the subsequent model, where batch represents the batch size, feature represents the number of features, and length represents the length of the generated data; According to the specific fault type or detection target, each time series segment is labeled accordingly; The labeled data is divided into training set, validation set and test set for subsequent model training.

7. A wind turbine fault detection method based on multivariate data and deep learning according to claim 1, characterized in that: Building a deep learning model based on state-space selection involves: The classical state space equation under continuous time conditions is: h′(t)=Ah(t)+Bx(t) (9) y(t)=Ch(t) (10) where h ′ (t) is the state change of the system in continuous time t, h(t) is the state vector of the system, A is the state matrix, B is the input matrix, C is the output matrix, x(t) and y(t) are the input and output vectors respectively; Discretization: the t =Ch t (12) in Calculated from A and B under the continuous system; Using convolution form: in, is the state vector in convolutional form; A, B, and C are all fixed parameters. A selection mechanism is added to make the B and C matrices change with the input. This is to increase the dimensions of the B and C matrices, incorporate the sequence length into B and C, and use the Linear layer for selection. This allows the B and C matrices to dynamically change each time the input is different.

8. A wind turbine fault detection method based on multivariate data and deep learning according to claim 1, characterized in that: The deep learning model based on state-space selection is a state-space algorithm that adds a selection mechanism to the main loop of the gated attention architecture. The state-space algorithm is regarded as a deep learning layer, and a convolutional layer is added before it to establish dependencies between different word vectors. A linear layer, a normalization layer, an activation layer, and a selection mechanism SSM block are added to the deep neural network.

9. A wind turbine fault detection method based on multivariate data and deep learning according to claim 1, characterized in that: The deep learning model based on state space selection adds residual connections to the selection mechanism SSM block and performs normalization before each input to make the model depth greater. The SSM blocks are stacked, and the number of stacking times is determined by the performance of the equipment performing the training. A normalization layer, a linear layer, and a softmax layer are added after the last block for fault classification and identification.

10. A wind turbine fault detection method based on multivariate data and deep learning according to any one of claims 1 to 9, characterized in that: Deploy the trained deep learning model locally or on the cloud, including: Select a set of local or cloud devices that can be used to deploy VAE models and selection mechanism SSM models, deploy the trained models to relevant devices, and perform real-time wind turbine fault detection.

Images