A wind turbine generator set small sample fault trend early warning method and system

Abstract

The present invention discloses a method and system for early warning of a small number of fault trends of a wind turbine, the method comprising: using a sensor to monitor the parameters of the wind turbine to form a parameter sample; formulating and identifying a fault; using an improved GAN model to generate a high-quality new sample similar to the original data based on the small number of parameter samples; enhancing the sample parameters and sending them to the improved TEMPO model for 4-hour-16-step fault prediction; using an extreme learning machine algorithm with prototype clustering preprocessing to perform fault identification on the output multivariate waveform, the system comprises a data acquisition module, a data enhancement and sample generation module, a fault prediction module, a data visualization and fault diagnosis module. The present invention improves the robustness of the network, helps to alleviate the problems of gradient vanishing and gradient explosion and accelerates the convergence speed of the neural network, and the introduction of the integrated normal cloud model strategy and the Cauchy mutation strategy updates and iterates the optimal solution of the algorithm, thereby improving the global search capability of the EDO algorithm.

Description

Technical Field

[0001] The present invention relates to the field of intelligent diagnosis of wind turbine faults, and in particular to a method and system for early warning of a small number of fault trends of wind turbines. Background Art

[0002] Wind turbine fault detection is a very important topic in the field of wind power generation, because wind turbines may face various faults and problems during operation, which not only affect the efficiency and reliability of wind power generation, but also may lead to safety risks and expensive maintenance costs. Therefore, timely and accurate detection and diagnosis of wind turbine faults is the key.

[0003] Fault prediction refers to the process of analyzing and processing historical data and then predicting future fault trends or remaining effective life through current data. With the development of technology and the needs of production and life, equipment maintenance has gradually evolved from accident maintenance to today's condition maintenance. Fault prediction technology is the basis for condition maintenance of equipment. Fault detection of wind turbines is the process of identifying faults or abnormal conditions in systems, equipment or machines. The goal of this process is to detect potential problems in a timely manner so that measures can be taken to avoid serious failures or losses. For wind turbines, fault detection involves monitoring and analyzing the operating status of wind turbines to identify anomalies or trends that may lead to failures. At present, although many advances have been made in wind turbine fault detection technology, there are still some disadvantages and challenges in practical applications. Wind turbines are equipped with many sensors and generate a large amount of data, which may be complex and difficult to manage. Summary of the invention

[0004] Purpose of the invention: The purpose of the present invention is to provide a method and system for early warning of fault trends of a small number of wind turbines.

[0005] Technical solution: The wind turbine generator set small sample fault trend early warning method of the present invention comprises the following steps:

[0006] (1) Use sensors to monitor the vibration, swing, sound pattern, temperature, voltage, current and oil temperature parameters of the wind turbine to form parameter samples;

[0007] (2) Based on the parameters sensed by the sensors, four types of faults are identified: mechanical fault, hydraulic system fault, electrical fault, and meteorological fault, and the fault module is identified;

[0008] (3) Based on the few parameter samples perceived by the sensor, the improved GAN model is used to generate high-quality new samples similar to the original data, thereby enhancing the data of the training set;

[0009] (4) The sample parameters are enhanced through the improved GAN model data to make them diverse parameter samples, and then sent to the improved TEMPO model for 4-hour-16-step fault prediction;

[0010] (5) The IMTBO algorithm includes the use of an elite reverse learning strategy to initialize individual climbers and expand the search range by calculating the reverse solution of the current solution; the use of an integrated normal cloud model strategy to replace the assisted climbing position update formula of the original climbing team algorithm to further improve the local development capability of the algorithm; and the use of a Cauchy mutation strategy to replace the random member replacement stage of the original climbing team optimization algorithm.

[0011] (6) Fault identification is performed on the multivariate waveform output from step (4) through the extreme learning machine algorithm with prototype clustering preprocessing, and a fault classification is formulated corresponding to step (2). Finally, the fault anomaly and results of the wind turbine are sent to the cloud data center.

[0012] Furthermore, the sensors used in step (1) include vibration sensors, temperature sensors, pressure sensors, current sensors, voltage sensors and rotation speed sensors.

[0013] Furthermore, the improved GAN model in step (3) is a combination of variational autoencoder VAE and deep fusion attention DFA; a two-stage model is used for data enhancement, namely, the variational autoencoder data enhancement stage and the deep attention mechanism optimization sample generation stage; both stages use a single-stage GAN as the backbone to regenerate the initial sample parameters with high quality.

[0014] Furthermore, the step (3) comprises:

[0015] (3.1) The improved GAN model, namely the VDGAN model, uses a two-stage model to generate samples, namely the variational autoencoder data enhancement stage and the deep attention mechanism optimization sample generation stage; both stages use a single-stage GAN as the backbone, and the initial samples generated in the first stage are input into the second stage to regenerate the initial samples with high quality;

[0016] (3.2) During the training process, the goal of the generator G is to generate as many real samples as possible to deceive the discriminator D. The final ideal state is D(G(z)) = 0.5, so the loss function formula of GAN is:

[0017]

[0018] Among them, V(D, G) represents the difference between the generated sample and the real sample, and the cross entropy loss function of the two classifications includes and Two parts: Indicates that the fixed generator G trains the discriminator D and updates the parameters of the discriminator D by maximizing the cross entropy loss V(D, G); Indicates that the fixed discriminator D trains the generator G;

[0019] (3.3) The encoding network in VAE consists of three fully connected layers. The first fully connected layer samples the input high-dimensional data, and the activation function uses Relu. The remaining two fully connected layers map the downsampled samples into two low-dimensional parameter features, namely the mean and variance of the normal distribution. The loss function of the encoding network is KL divergence, which is used to measure the difference between the implicit vector distribution and the standard normal distribution. The KL loss function is:

[0020]

[0021] Among them, μ and σ 2 are the mean and variance of the encoding network output respectively; F KL Indicates the calculation of the KL divergence of two approximate distributions, which is used here to represent the KL divergence between the data distribution sought by the present invention and the standard normal distribution N(0,1);

[0022] (3.4) The high-quality new samples obtained through the first stage are input into the second stage to obtain the final optimized generated image;

[0023] The input of the second stage is the samples and feature vectors generated in the first stage. The encoder is used to remove the multi-layer perceptron classification head in the original network model, and then the vector originally used for additional classification prediction output results is removed as the encoder of the generator at this stage to obtain the image feature Y, which is expressed as:

[0024]

[0025] y l ′=MSA(LN(y l-1 ))+y l-1 ; l = 1, 2, ..., L.

[0026] y l =MLP(LN(y l ′))+y l ′; l = 1, 2, ..., L.

[0027] Y=y1+y2+y3+...+y L

[0028] in: is the trainable linear projection parameter, E pos ∈R (N+1)×Kis the position encoding, LN(·) is the layer normalization function; the ViT encoder consists of L Transformer encoder blocks, and the input y of each encoder block is l-1 It is the output of the previous encoder block. The output y of the current encoder block is obtained through multi-head self-attention and MLP calculation l After obtaining the sample feature Y, the feature vector Splice it into the image features to get a new feature vector Y′, which is expressed as:

[0029]

[0030] The feature vector Y′ is input into a series of convolution decoding to obtain the second-stage generated samples, and then the discriminator identifies the authenticity of the samples to supervise the quality of image generation; the loss function of the second stage is composed of the generator loss function and the discriminator loss function as follows:

[0031]

[0032]

[0033] in: is the second stage generator loss function, is the second stage discriminator loss function, is the Deep Attention Multimodal Similarity Model (DAMSM) loss function, is the feature matching loss function, λ1 and λ2 are the corresponding weight parameters of DAMSM loss and feature matching loss in the second stage generator respectively.

[0034] Furthermore, in the step (3.3), the decoding network is composed of two layers of fully connected networks. The parameter settings of the two layers of networks are opposite to those of the encoding network. The input low-dimensional hidden variables are feature-enlarged to finally generate generated data of the same size as the input data. The loss function is:

[0035]

[0036] The discriminant network consists of 2 convolutional layers and 3 fully connected layers. The activation function between each network layer uses Relu. The output layer outputs the true and false labels through the Sigmoid activation function. The loss function of the discriminant network is:

[0037]

[0038] Among them, x~P data (x) indicates that the real sample obeys P data (x) distribution; z~P z (z) indicates the expectation to obey P z (z) distribution.

[0039] Furthermore, the improved TEMPO model in step (4) is a combination of variational mode decomposition and an improved mountaineering optimization algorithm; VMD is used to replace the original signal decomposition module of TEMPO, and the IMTBO algorithm is used to optimize the parameters of TEMPO.

[0040] Furthermore, the step (4) comprises:

[0041] (4.1) Problem definition: Given the observed values ​​of the previous K timestamps, the task of multivariate time series prediction is to predict the value of the next H timestamps. The formula is:

[0042]

[0043] in, is the estimated vector of the Hth step of timestamp t of channel f corresponding to the fth feature; historical value The model F, the parameter φ and the prompt V i to infer;

[0044] (4.2) Put the three parameters of trend, seasonality and residual into the improved mountaineering team optimization algorithm for optimization and then proceed to the subsequent steps;

[0045] (4.3) Time series input: Given input X∈R n×L , where n is the feature channel size and L is the length of the time series, it can be expressed as:

[0046]

[0047] Where f is the channel index of the multivariate time series input, and trend captures the underlying long-term patterns in the data, where m = 2k + 1 and k is the average step size; the seasonal component X s ∈R n×L Includes repeated short-term cycles, estimated after removing the trend component; the residual component X R ∈R n×L It represents the remaining part of the data after extracting the trend and seasonality; then the variational mode decomposition VMD is introduced to replace the data decomposition module of the original model;

[0048] (4.4) Prompt design: First, the trend-specific prompts are converted into word embedding space, and then linearly transformed to obtain a learnable trend prompt vector V t ; The combination of this prompt and the time series representation is embedded as:

[0049] x T =[V t ;P T ]

[0050] Among them, x T represents the set of embeddings along the time axis; this concatenation process reflects the seasonal and residual components, producing X S and X R ;

[0051] (4.5) GPT architecture: Use the decoder-based GPT as the backbone to build the foundation of time series representation, connect the prompts and different components together, and put them into the GPT block; specifically, the input of the time series embedding in this architecture is expressed as:

[0052]

[0053] in, Corresponding to the connection operation, x * Indicates different parameters.

[0054] Furthermore, the step (5) comprises:

[0055] (5.1) Set the relevant parameters of the IMTBO algorithm: population size, dimension size, upper and lower bounds of the path search space;

[0056] (5.2) The elite reverse learning strategy is used to replace the original random initialization in the IMTBO algorithm. By calculating the reverse solution of the current solution, the diversity of the population is increased, and the optimal solution is selected from the current solution and the reverse solution as the new generation of individuals;

[0057] The elite reverse solution is defined as: Assume that the extreme point corresponding to the individual in the population is the elite individual, that is Its reverse solution It is expressed as:

[0058]

[0059] Where K is the dynamic coefficient on (0, 1); α j =min(X i，j ), β j =max(X i，j ), α j , β j is a dynamic boundary. If the dynamic boundary operation is If the solution exceeds the boundary and becomes infeasible, it can be reset by using a random generation method. The reset method is as follows:

[0060]

[0061] (5.3) The cloud model is described by three parameters: expectation Ex, entropy En, and super entropy He. When En increases, the cloud droplet distribution range becomes larger. When He increases, the discreteness of the cloud droplets increases synchronously, which reflects the randomness and fuzziness of the cloud droplet distribution. The normal cloud droplet generation process can be defined as follows:

[0062] X = [x1, x2, ..., x Nd ]=Gnc(Ex,En,He,Nd)

[0063] Where Nd is the expected number of cloud droplets; En is used to control the remaining position solutions, and He is used to adjust the position discreteness. The formula is as follows:

[0064] Position=Gnc(Position best ,En,He,dim)

[0065]

[0066] He=En×10 -ξ

[0067] Among them, Position best is the optimal position of the current population; dim is the dimension set by the algorithm; λ = 0.3, τ = 2, ξ = 2; t is the current number of iterations, and T is the maximum number of iterations;

[0068] (5.4) In the algorithm, in order to avoid falling into the local optimum and move towards the optimal solution position, the position update formula of the solution at this stage is:

[0069]

[0070] in, is the new position obtained, X Avalanche is the position of other solutions when the local optimal state occurs randomly;

[0071] (5.5) When a local optimum occurs, the entire algorithm will try its best to save each solution from being trapped. The position update formula is as follows:

[0072]

[0073] Among them, X Team is the average position of all solutions;

[0074] (5.6) The position update formula using the Cauchy mutation operator is:

[0075]

[0076] Among them, cauchy(0,1) is the standard Cauchy distribution function; the one-dimensional Cauchy variogram centered at the origin is as follows:

[0077]

[0078] Cauchy mutation is used to perturb the individuals in the position update of the climbers and expand the search scale of the climbing team optimization algorithm.

[0079] Furthermore, the step (6) comprises:

[0080] Assume that there is a data set φ=((x;t i ), x i ∈R n , t i ∈R m ) N , where x i =[x i1 , x i2 , ..., x in ] T is the network input, t i =[t i1 , t i2 , ..., t in ] T is the network output; the network structure of ELM can be described as:

[0081]

[0082] Where G(α i , b i , x) is the output of the i-th hidden layer node; the minimum error is constructed as ||HT||, and by solving the least squares of H=T, we can get β=(H T H) -1 H T T, where: H is the hidden layer node output; T is the label. Based on the RKELM algorithm of prototype clustering preprocessing, the network has more hidden layer nodes.

[0083] The wind turbine generator set small sample fault trend warning system includes a data acquisition module, a data enhancement and sample generation module, a fault prediction module, a data visualization and fault diagnosis module;

[0084] The data acquisition module uses a variety of sensors to collect data from various components of the wind turbine in real time, and formulates four fault categories;

[0085] The data enhancement and sample generation module uses the VAE-DFA-GAN model to generate high-quality original data based on the few sample parameters collected by the data collection module;

[0086] The fault prediction module uses the VMD-TEMPO-IMTBO model to predict faults on the data of the above modules and generates a 16-step prediction waveform with one point every 15 minutes for 4 hours;

[0087] The data visualization and fault diagnosis module identifies the fault condition of the wind turbine and the 4-hour 16-step prediction waveform through the extreme learning machine algorithm with prototype clustering preprocessing, and finally displays it in real time through cloud platform data visualization.

[0088] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0089] The addition of elite reverse learning strategy improves the robustness of the network, helps alleviate the problems of gradient vanishing and gradient exploding, and accelerates the convergence speed of the neural network. The introduction of integrated normal cloud model strategy and Cauchy mutation strategy updates and iterates the optimal solution of the algorithm, improving the global search capability of the EDO algorithm.

[0090] The extreme learning machine algorithm of prototype clustering preprocessing can reduce the dimensionality of data through the prototype clustering method, simplifying the data into a set of representative prototype data, which not only reduces the data dimension but also reduces the noise, thereby improving the stability and performance of the model; through prototype clustering, the size of the data set can be reduced and more representative prototype data can be selected, which helps to reduce the computational burden of ELM during training and prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] Figure 1 is a flow chart of the present invention;

[0092] Figure 2 Schematic diagram of the improved VDGAN model architecture;

[0093] Figure 3 This is a schematic diagram of the improved TEMPO model architecture;

[0094] Figure 4 Flowchart of RKELM for decision-level fusion based on prototype clustering preprocessing. DETAILED DESCRIPTION

[0095] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0096] like Figure 1 As shown, the wind turbine generator set small sample fault trend early warning method of the present invention specifically includes the following steps:

[0097] Step 1: Select appropriate sensors according to the parameters that need to be monitored; vibration sensors are placed in the gearbox, main shaft, tower base and tower body; pressure sensors are used to monitor oil pressure or air pressure, etc.; current and voltage sensors are placed in the core position of the generator to monitor its charging and discharging status; speed sensors are placed on the wind turbine blades to monitor the dynamic balance and normal swing of the blades; then use edge computing devices to process and analyze the on-site situation, and transmit the data of each parameter to the central server in real time for subsequent processing when necessary.

[0098] Step 2: According to the steps described in step 1 and the data parameters obtained, the present invention has formulated four types of faults: mechanical faults caused by vibration and swing abnormalities caused by gear boxes, main shafts, tower bases and tower bodies, hydraulic system faults caused by pressure abnormalities caused by oil pressure and air pressure, electrical faults caused by current and voltage abnormalities caused by generators, and meteorological faults caused by atmospheric temperature.

[0099] Step 3: In actual work, the wind turbine bearings are in a complex environment, and it is difficult to obtain a large amount of fault data. The serious imbalance between normal data and fault data causes insufficient training of the deep model and low diagnostic accuracy. Therefore, the present invention proposes a generative adversarial network (VDGAN) model combining a variational autoencoder and a deep attention mechanism for few sample parameters, such as Figure 2 As shown, the specific implementation steps are:

[0100] Step 3.1: The VDGAN model proposed in the present invention uses a two-stage model to generate samples, namely, the variational autoencoder data enhancement stage and the deep attention mechanism optimization sample generation stage; both stages use a single-stage GAN as the backbone, and the initial samples generated in the first stage are input into the second stage to regenerate the initial samples with high quality;

[0101] Step 3.2: During the training process, the goal of the generator G is to generate real samples as much as possible to deceive the discriminator D. The goal of D is to distinguish the samples generated by G from the real samples as much as possible. The final ideal state is D(G(z)) = 0.5, so the loss function formula of GAN is:

[0102]

[0103] The left side of the formula V(D, G) represents the difference between the generated sample and the real sample. The cross entropy loss function of the binary classification (true and false) includes and Two parts: Indicates that the fixed generator G trains the discriminator D and updates the parameters of the discriminator D by maximizing the cross entropy loss V(D, G); Indicates that the fixed discriminator D is used to train the generator G. The generator should minimize the cross entropy function while maximizing the cross entropy loss V(D, G) between true and false samples in the discriminator.

[0104] Step 3.3: The encoding network in VAE consists of three fully connected layers. The first fully connected layer samples the input high-dimensional data, and the activation function uses Relu. The remaining two fully connected layers map the downsampled samples into two low-dimensional parameter features, namely the mean and variance of the normal distribution. The loss function of the encoding network is KL divergence, which is used to measure the difference between the implicit vector distribution and the standard normal distribution. The KL loss function is:

[0105]

[0106] In the formula, μ and σ 2 are the mean and variance of the encoding network output respectively; Formula (2) represents the calculation of the KL divergence of two approximate distributions, which is used here to represent the KL divergence between the data distribution sought by the present invention and the standard normal distribution N(0,1);

[0107] The decoding network (generating network) consists of two layers of fully connected networks. The parameter settings of these two layers are opposite to those of the encoding network. The input low-dimensional hidden variables are feature-enlarged to generate generated data of the same size as the input data. The Relu activation function is used between the two network layers, and the final output layer uses the Sigmoid activation function. The optimization goal of the generating network is to minimize the feature loss between the generated samples and the real samples, and to ensure that the generated data can deceive the discriminant network. Its loss function is

[0108]

[0109] The discriminant network consists of 2 convolutional layers and 3 fully connected layers. The activation function between each network layer uses Relu. The output layer outputs the true and false labels through the Sigmoid activation function. The loss function of the discriminant network is:

[0110]

[0111] In the formula, x~P data (x) indicates that the real sample obeys P data (x) distribution; z~P z (z) indicates the expectation to obey P z (z) distribution;

[0112] Step 3.4: The high-quality new samples obtained through the first stage are input into the second stage (deep attention mechanism optimization sample generation stage) to obtain the final optimized generated image;

[0113] The input of the second stage is the samples and feature vectors generated in the first stage. The encoder (ViT) is used to remove the multi-layer perceptron classification head in the original network model, and then the vector originally used for additional classification prediction output results is removed as the encoder of the generator at this stage to obtain the image feature Y, which is expressed as

[0114]

[0115] y l ′=MSA(LN(y l-1 ))+y l-1 ; l = 1, 2, ..., L. (6)

[0116] y l =MLP(LN(y l ′))+y l ′; l=1, 2,..., L. (7)

[0117] Y=y1+y2+y3+...+y L (8)

[0118] Where: is the trainable linear projection parameter, E pos ∈R (N+1)×K is the position encoding, LN(·) is the layer normalization function; the ViT encoder consists of L Transformer encoder blocks, and the input y of each encoder block is l-1 , which is the output of the previous encoder block, is calculated by multi-head self-attention (MSA) and MLP to obtain the output y of the current encoder block l After obtaining the sample feature Y, the feature vector Splice it into the image features to get a new feature vector Y′, which is expressed as:

[0119]

[0120] The feature vector Y′ is input into a series of convolution decoding to obtain the second-stage generated samples, which are then identified by the discriminator to monitor the quality of image generation.

[0121] The generator and discriminator are trained alternately, and the generator loss and discriminator loss in the data enhancement stage are optimized through the variational autoencoder data enhancement stage and the deep attention mechanism to obtain samples that meet the text conditions; the loss function of the DFA-GAN model is the second-stage loss function, which is composed of the generator loss function and the discriminator loss function;

[0122]

[0123] Where: is the second stage generator loss function, is the second stage discriminator loss function, is the Deep Attention Multimodal Similarity Model (DAMSM) loss function, is the feature matching loss function, λ1 and λ2 are the corresponding weight parameters of the DAMSM loss and feature matching loss in the second stage generator respectively;

[0124]

[0125] In the second stage the generator is added The purpose is to avoid the instability of the adversarial generative network during training, so that the generator can better capture the similarities between the feature representations of the intermediate layers and help the generator generate more realistic and coherent images.

[0126] Step 4: Use the improved TEMPO model to perform 4-hour-16-step fault prediction, as shown in Figure 3 As shown, the specific implementation steps are:

[0127] Step 4.1: Problem definition: Given the observed values ​​of the previous K timestamps, the task of multivariate time series prediction is to predict the value of the next H timestamps. The formula is:

[0128]

[0129] In the formula, is the estimated vector of the Hth step of timestamp t of channel f corresponding to the fth feature; historical value The model F, the parameter φ and the prompt V i To infer; considering that the basic model has strong generalization ability on unseen data sets, the present invention uses the zero-shot learning configuration by default without specific instructions; this method requires that the model cannot know the historical value and level value of the target data set during the training process;

[0130] Step 4.2: Put the three parameters of trend, seasonality and residual in step 4.3 into the improved mountaineering team optimization algorithm for optimization and then proceed to the subsequent steps;

[0131] Step 4.3: Time series input: For time series data, representing complex input by decomposing it into meaningful components (such as trend and seasonal components) can help to optimally extract information; in this invention, given an input X∈R n ×L , where n is the feature channel size and L is the length of the time series, it can be expressed as:

[0132]

[0133] Where f is the channel index of the multivariate time series input, trend captures the underlying long-term patterns in the data, where m = 2k + 1 and k is the average step size; the seasonal component X s ∈R n×L Includes repeated short-term cycles, which can be estimated after removing the trend component; the residual component X R ∈R n×L It represents the remaining part of the data after extracting the trend and seasonality; then the variational mode decomposition (VMD) is introduced to replace the data decomposition module of the original model:

[0134] Assume that the initial signal f is composed of modal components u k (f) composition, first calculate u k The unilateral spectrum of (f) is Hilbert transformed and then passes through the center frequency w k , modulation u k (f) spectrum to baseband, and finally solve for u k The bandwidth of (f) corresponds to the constraint condition as shown in formula (15):

[0135]

[0136] Where, δ(f): Dirac function, Gradient function, *: convolution operation

[0137] The following is the variational model calculation process:

[0138] Step 4.3.1: First, introduce the penalty factor α and the Lagrangian operator λ(f) in equation (16), so that equation (16) becomes a non-constrained problem. The transformed expression is as follows:

[0139]

[0140] Where, <>: inner product operation;

[0141] Step 4.3.2: Based on formula (16), for u k , w k and λ are repeatedly iterated and modified to find the final optimal solution; find Expression in frequency domain Indicates the cumulative number of iterations that have occurred;

[0142] Will and Perform initialization processing and update according to the expression of formulas (17) to (19)

[0143]

[0144] The final stopping condition of the iteration is:

[0145]

[0146] In the formula, ε is the discrimination accuracy and is greater than 0;

[0147] Step 4.4: Prompt Design: This approach involves generating different prompts corresponding to each major time series component: trend, seasonality, and residual; predicting future time steps based on trend, seasonality, and residual is a template for deriving component-specific prompts in the present invention; trend-specific prompts are first converted into word embedding space and then linearly transformed to derive a learnable trend prompt vector V t ; The combination of this prompt and the time series representation is embedded as:

[0148] x T =[V t ;P T ] (twenty one)

[0149] In the formula, x T represents the set of embeddings along the time axis; this concatenation process reflects the seasonal and residual components, producing X S and X R ;

[0150] Step 4.5: GPT Architecture:

[0151] This architecture uses the decoder-based GPT as the backbone to build the foundation of time series representation. In order to effectively utilize the decomposed semantic information, this architecture connects the prompts and different components together and puts them into the GPT block; specifically, the input of the time series embedding in this architecture can be expressed as:

[0152]

[0153] in, Corresponding to the connection operation, x * represents different parameters; in the GPT block, we adopt the strategy of selectively updating the position embedding layer and the normalization layer gradient. In addition, low-rank adaptation is also used to effectively adapt to the changing time series distribution because it can be adapted with fewer parameters;

[0154] The overall prediction result is the additive combination of the predictions of each component. Finally, the output Z of the n features of the GPT block can be divided into (outputs corresponding to trend, seasonality, residuals), and then each Z component is fed into a fully connected layer to generate a forecast Where L His the prediction length, and the prediction result can be expressed as: Afterwards, the architecture denormalizes Y according to the corresponding statistics used in the normalization step: By recombining these additional elements, the approach of the present architecture aims to reconstruct the complete temporal trajectory that best represents the underlying dynamics at different time scales captured by the decomposed input representation;

[0155] To achieve interpretability, this architecture explores the linear and nonlinear interactions between the contributions of trend, seasonality, and residual components to the final output. Therefore, an interpretable generalized additive model (GAM) is constructed based on the output of GPT to understand how the three components interact: In the formula, F φ is a normalizing constant, the footnote f corresponds to the trend, seasonal and residual components, I t is a set of multiple interacting components, then, a first-order sensitivity index can be calculated to measure the sensitivity of each component.

[0156] Step 5: The specific implementation steps of optimizing the parameters in the TEMPO model using the improved IMTBO algorithm are as follows:

[0157] Step 5.1: Set the relevant parameters of the IMTBO algorithm: population size, dimension size, upper and lower bounds of the path search space;

[0158] Step 5.2: Use the elite reverse learning strategy to replace the original random initialization in the IMTBO algorithm, increase the diversity of the population by calculating the reverse solution of the current solution, and select the optimal solution from the current solution and the reverse solution as the new generation of individuals; the elite reverse solution is defined as: assuming that the extreme point corresponding to the individual in the population is the elite individual, that is:

[0159] Its reverse solution It can be expressed as:

[0160]

[0161] Where K is the dynamic coefficient on (0, 1); α j =min(X i，j ), β j =max(X i，j ), α j , β j The dynamic boundary overcomes the disadvantage of the fixed boundary that it is difficult to save the search experience, so that the elite reverse solution can be searched in a narrow space and is not easy to fall into the local optimum; if the dynamic boundary operation makes If the solution exceeds the boundary and becomes infeasible, it can be reset by using a random generation method. The reset method is as follows:

[0162]

[0163] Step 5.3: The integrated normal cloud model strategy is used to replace the assisting climbing position update formula in the original climbing team algorithm, which further improves the local development ability of the algorithm and greatly reduces the probability of the algorithm falling into a local solution; the cloud model is described by three parameters: expectation Ex, entropy En, and super entropy He. When En increases, the distribution range of cloud droplets increases accordingly. When He increases, the discrete degree of cloud droplets increases synchronously, which indirectly reflects the randomness and fuzziness of cloud droplet distribution. The forward normal cloud generator is an algorithm that generates cloud droplets that basically obey the normal distribution. Cloud droplets are generated by setting parameters until the desired number of cloud droplets is generated. The normal cloud droplet generation process can be defined as follows:

[0164] X = [x1, x2, ..., x Nd ]=Gnc(Ex,En,He,Nd) (25)

[0165] Where Nd is the expected number of cloud droplets. The normal cloud model (26) is introduced as a new updating mechanism to assist mountaineering position in the IMTBO algorithm. The optimal position solution is developed by the expected value Ex of the normal cloud model, the remaining position solutions are regulated by En, and the position discreteness is adjusted by He. The formula is as follows:

[0166] Position=Gnc(Position best ,En,He,dim) (26)

[0167]

[0168] He=En×10 -ξ (28)

[0169] Among them, Position best is the optimal position of the current population; dim is the dimension set by the algorithm; λ = 0.3, τ = 2, ξ = 2; t is the current number of iterations, and T is the maximum number of iterations;

[0170] Step 5.4: In the disaster threat stage of the IMTBO algorithm, the occurrence of avalanches or falling off cliffs may threaten the lives of the climbers, thereby preventing the climbers from reaching the top of the mountain. In the present invention, this forces the algorithm to fall into a local optimal state. In the algorithm, in order to avoid falling into a local optimal state and move toward the optimal solution position, the position update formula of the solution in this stage is:

[0171]

[0172] In the formula, is the new position obtained, XAvalanche is the position of other solutions when the local optimal state occurs randomly; other parameters have the same meaning as above;

[0173] Step 5.5: During the solution process, when a local optimum occurs, the entire algorithm will try its best to save each solution from falling into the position update formula as follows:

[0174]

[0175] Where, X Team is the average position of all solutions; other parameters have the same meaning as above;

[0176] Step 5.6: Use the Cauchy mutation strategy to replace the random member replacement phase of the original mountaineering team optimization algorithm. In the random replacement solution phase, a new solution is generated by using a normal distribution random number. The position update formula of the Cauchy mutation operator used in this phase is:

[0177]

[0178] Where cauchy(0,1) is the standard Cauchy distribution function; the one-dimensional Cauchy variogram centered at the origin is as follows:

[0179]

[0180] The Cauchy distribution is similar to the standard normal distribution in that it is a continuous probability distribution with a smaller value at the origin, more elongated values ​​at both ends, and a slower rate of approaching zero. Therefore, it can produce greater disturbances than the normal distribution. Therefore, the Cauchy mutation is used to perturb the individuals in the position update of the climbers, thereby expanding the search scale of the climbers' optimization algorithm and improving the algorithm's ability to escape from the local optimum.

[0181] Step 6: Use the improved ELM algorithm to perform fault identification on the multivariate waveform output from step 4, such as Figure 4 As shown, the specific fact steps are as follows:

[0182] Step 6.1: Assume that there is a data set φ = ((x; t i ), x i ∈R n , t i ∈R m ) N , where x i =[x i1 , x i2 , ..., x in ] T is the network input, t i =[t i1 , t i2 , ..., t in ]T is the network output; the network structure of ELM can be described as:

[0183]

[0184] In the formula, G(α i , b i , x) is the output of the i-th hidden layer node; the minimum error is constructed as ||HT||, and by solving the least squares of H=T, we can get β=(H T H) -1 H T T, where: H is the hidden layer node output; T is the label;

[0185] Step 6.2: When there are fewer hidden nodes in the ELM network, some support vectors may have greater correlation, resulting in waste of hidden nodes and unstable recognition. The present invention proposes an RKELM algorithm based on prototype clustering preprocessing, which effectively solves the above problem;

[0186] Assume that the training sample is X = [x1, x2, ..., x N1 ], contains N1 samples from C-type targets, and the number of hidden nodes in RKELM is N h , then first perform prototype clustering on X and divide it into N h disjoint training sample clusters, the resulting prototype vector set is Y P =[y1,y2,...,y Nh ],y j is the prototype vector of the jth cluster, j = 1, 2, ..., N h ; The clustering structure of the sample can be represented by the prototype vector set Y P Characterize, and Y P The distance between the two vectors in the P As the support vector of RKELM, it can more comprehensively characterize the feature space of the target and improve the target recognition performance. Without loss of generality, the prototype clustering algorithm used in the present invention is K-means++;

[0187] set up is the category matrix of X, where t n is the nth training sample x n The category vector of x n If the category is c, then t n The cth element in is 1, the remaining elements are 0, n = 1, 2, ..., N1, c = 1, 2, ..., C. The training phase of the RKELM algorithm based on prototype clustering preprocessing is carried out according to the following steps:

[0188] Step 6.2.1: Use K-means++ algorithm to cluster the training sample X, where the number of clusters is N. h , the cluster center is Y P =[y1,y2,...,y Nh ];

[0189] Step 6.2.2: Set the cluster center Y P As the hidden layer node of RKELM, calculate the kernel matrix K0:

[0190]

[0191] Among them, K(·,·) is the kernel function;

[0192] Step 6.2.3: Calculate the output layer weight matrix according to the empirical error minimization criterion

[0193]

[0194] Among them, ξ is the regularization parameter, I is N h dimensional identity matrix; in the test phase, let x u is a test sample, then its category vector can be calculated as:

[0195] t u =K(x u , Y p )β (36)

[0196] Where, t u The category corresponding to the maximum value is the sample x u Category.

[0197] Based on the same inventive concept, the present invention also proposes a wind turbine generator set small sample fault trend warning method, including a data acquisition module, a data enhancement and sample generation module, a fault prediction module, a data visualization and fault diagnosis module; the data acquisition module uses a variety of sensors to perform real-time data collection on various components of the wind turbine generator set; and formulates four fault categories; the data enhancement and sample generation module uses the VAE-DFA-GAN model to generate high-quality original data for the small sample parameters collected by the data acquisition module; the fault prediction module uses the VMD-TEMPO-IMTBO model to predict faults for the data of the above modules, and generates a 16-step prediction waveform of 4 hours and 1 point every 15 minutes; the data visualization and fault diagnosis module identifies faults of the 4-hour 16-step prediction waveform for the fault condition of the wind turbine generator set and through the extreme learning machine (RKELM) algorithm with prototype clustering preprocessing, and finally displays it in real time through cloud platform data visualization.

Claims

1. A wind turbine generator set small sample fault trend early warning method, characterized in that: The steps include: (1) Use sensors to monitor the vibration, swing, sound pattern, temperature, voltage, current and oil temperature parameters of the wind turbine to form parameter samples; (2) Based on the parameters sensed by the sensors, four types of faults are identified: mechanical fault, hydraulic system fault, electrical fault, and meteorological fault, and the fault module is identified; (3) Based on the few parameter samples perceived by the sensor, the improved GAN model is used to generate high-quality new samples similar to the original data, thereby enhancing the data of the training set; (4) The sample parameters are enhanced through the improved GAN model data to make them diverse parameter samples, and then sent to the improved TEMPO model for 4-hour-16-step fault prediction; (5) The IMTBO algorithm includes using an elite reverse learning strategy to initialize individual climbers and expand the search range by calculating the reverse solution of the current solution; using an integrated normal cloud model strategy to replace the auxiliary climbing position update formula of the original climbing team algorithm to further improve the local development ability of the algorithm; using the Cauchy mutation strategy to replace the random member replacement stage of the original climbing team optimization algorithm; (6) Fault identification is performed on the multivariate waveform output from step (4) through the extreme learning machine algorithm with prototype clustering preprocessing, and a fault classification is formulated corresponding to step (2). Finally, the fault anomaly and results of the wind turbine are sent to the cloud data center.

2. The wind turbine generator set small sample fault trend early warning method according to claim 1 is characterized in that: The sensors used in step (1) include vibration sensors, temperature sensors, pressure sensors, current sensors, voltage sensors and rotation speed sensors.

3. The wind turbine generator set small sample fault trend early warning method according to claim 1, characterized in that: The improved GAN model in step (3) is a combination of variational autoencoder VAE and deep fusion attention DFA; a two-stage model is used for data enhancement, namely, variational autoencoder data enhancement stage and deep attention mechanism optimization sample generation stage; both stages use a single-stage GAN as the backbone to regenerate the initial sample parameters with high quality.

4. The wind turbine generator set small sample fault trend early warning method according to claim 1, characterized in that: The step (3) comprises: (3.1) The improved GAN model, namely the VDGAN model, uses a two-stage model to generate samples, namely the variational autoencoder data enhancement stage and the deep attention mechanism optimization sample generation stage; both stages use a single-stage GAN as the backbone, and the initial samples generated in the first stage are input into the second stage to regenerate the initial samples with high quality; (3.2) During the training process, the goal of the generator G is to generate as many real samples as possible to deceive the discriminator D. The final ideal state is D(G(z)) = 0.5, so the loss function formula of GAN is: Among them, V(D, G) represents the difference between the generated sample and the real sample, and the cross entropy loss function of the two classifications includes and Two parts: Indicates that the fixed generator G trains the discriminator D and updates the parameters of the discriminator D by maximizing the cross entropy loss V(D, G); Indicates that the fixed discriminator D trains the generator G; (3.3) The encoding network in VAE consists of three fully connected layers. The first fully connected layer samples the input high-dimensional data, and the activation function uses Relu. The remaining two fully connected layers map the downsampled samples into two low-dimensional parameter features, namely the mean and variance of the normal distribution. The loss function of the encoding network is KL divergence, which is used to measure the difference between the implicit vector distribution and the standard normal distribution. The KL loss function is: Among them, μ and σ 2 are the mean and variance of the encoding network output respectively; F KL Indicates the calculation of the KL divergence of two approximate distributions, which is used here to represent the KL divergence between the data distribution sought by the present invention and the standard normal distribution N(0,1); (3.4) The high-quality new samples obtained through the first stage are input into the second stage to obtain the final optimized generated image; The input of the second stage is the samples and feature vectors generated in the first stage. The encoder is used to remove the multi-layer perceptron classification head in the original network model, and then the vector originally used for additional classification prediction output results is removed as the encoder of the generator at this stage to obtain the image feature Y, which is expressed as: y l ′=MSA(LN(y l-1 ))+y l-1 ;l=1,2,...,L. and l =MLP(LN(and l ′))+y l ′;l=1,2,...,L. <h2 style=";text-align:left;direction:ltr">Y = y1 + y2 + y3 +... + y<h2 style=";text-align:left;direction:ltr"> L in: is the trainable linear projection parameter, E pos ∈R (N+1)×K is the position encoding, LN(·) is the layer normalization function; the ViT encoder consists of L Transformer encoder blocks, and the input y of each encoder block is l-1 It is the output of the previous encoder block. The output y of the current encoder block is obtained through multi-head self-attention and MLP calculation l After obtaining the sample feature Y, the feature vector Spliced ​​into the image features, a new feature vector Y′ is obtained, which is expressed as: The feature vector Y′ is input into a series of convolution decoding to obtain the second-stage generated samples, and then the discriminator identifies the authenticity of the samples to supervise the quality of image generation; the loss function of the second stage is composed of the generator loss function and the discriminator loss function as follows: in: is the second stage generator loss function, is the second stage discriminator loss function, is the Deep Attention Multimodal Similarity Model (DAMSM) loss function, is the feature matching loss function, λ1 and λ2 are the corresponding weight parameters of DAMSM loss and feature matching loss in the second stage generator respectively.

5. The wind turbine generator set small sample fault trend early warning method according to claim 4 is characterized in that: In the step (3.3), the decoding network is composed of two layers of fully connected networks. The parameter settings of the two layers of networks are opposite to those of the encoding network. The input low-dimensional hidden variables are feature-enlarged to finally generate generated data of the same size as the input data. The loss function is: The discriminant network consists of 2 convolutional layers and 3 fully connected layers. The activation function between each network layer uses Relu. The output layer outputs the true and false labels through the Sigmoid activation function. The loss function of the discriminant network is: Among them, x~P data (x) indicates that the real sample obeys P data (x) distribution; z~P z (z) indicates the expectation to obey P z (z) distribution.

6. The wind turbine generator set small sample fault trend early warning method according to claim 1, characterized in that: The improved TEMPO model in step (4) is a combination of variational mode decomposition and an improved mountaineering team optimization algorithm; VMD was used to replace the original signal decomposition module of TEMPO, and the IMTBO algorithm was used to optimize the parameters of TEMPO.

7. The wind turbine generator set small sample fault trend early warning method according to claim 1, characterized in that: The step (4) comprises: (4.1) Problem definition: Given the observed values ​​of the previous K timestamps, the task of multivariate time series prediction is to predict the value of the next H timestamps. The formula is: in, is the estimated vector of the Hth step of timestamp t of channel f corresponding to the fth feature; historical value The model F, the parameter φ and the prompt V i to infer; (4.2) Put the three parameters of trend, seasonality and residual into the improved mountaineering team optimization algorithm for optimization and then proceed to the subsequent steps; (4.3) Time series input: Given input X∈R n×L , where n is the feature channel size and L is the length of the time series, it can be expressed as: Where f is the channel index of the multivariate time series input, and trend captures the underlying long-term patterns in the data, where m = 2k + 1 and k is the average step size; the seasonal component X s ∈R n×L Includes repeated short-term cycles, estimated after removing the trend component; the residual component X R ∈R n×L It represents the remaining part of the data after extracting the trend and seasonality; then the variational mode decomposition VMD is introduced to replace the data decomposition module of the original model; (4.4) Prompt design: First, the trend-specific prompts are converted into word embedding space, and then linearly transformed to obtain a learnable trend prompt vector V t ; The combination of this prompt and the time series representation is embedded as: x T =[V t ;P T ] Among them, x T represents the set of embeddings along the time axis; this concatenation process reflects the seasonal and residual components, producing X S and X R ; (4.5) GPT architecture: Use the decoder-based GPT as the backbone to build the foundation of time series representation, connect the prompts and different components together, and put them into the GPT block; specifically, the input of the time series embedding in this architecture is expressed as: in, Corresponding to the connection operation, x * Indicates different parameters.

8. The wind turbine generator set small sample fault trend early warning method according to claim 1, characterized in that: The step (5) comprises: (5.1) Set the relevant parameters of the IMTBO algorithm: population size, dimension size, upper and lower bounds of the path search space; (5.2) The elite reverse learning strategy is used to replace the original random initialization in the IMTBO algorithm. By calculating the reverse solution of the current solution, the diversity of the population is increased, and the optimal solution is selected from the current solution and the reverse solution as the new generation of individuals; The elite reverse solution is defined as: Assume that the extreme point corresponding to the individual in the population is the elite individual, that is Its reverse solution It is expressed as: Where K is the dynamic coefficient on (0, 1); α j =min(X i，j ), β j =max(X i，j ), α j , β j is a dynamic boundary. If the dynamic boundary operation is If the solution exceeds the boundary and becomes infeasible, it can be reset by using a random generation method. The reset method is as follows: (5.3) The cloud model is described by three parameters: expectation Ex, entropy En, and super entropy He. When En increases, the cloud droplet distribution range becomes larger. When He increases, the discreteness of the cloud droplets increases synchronously, which reflects the randomness and fuzziness of the cloud droplet distribution. The normal cloud droplet generation process can be defined as follows: X=[x1,x2,...,x Nd ]=Gnc(Ex,E,He,Nd) Where Nd is the expected number of cloud droplets; En is used to control the remaining position solutions, and He is used to adjust the position discreteness. The formula is as follows: Position=Gnc(Position best ,En,He,dim) He=En×10 -ξ Among them, Position best is the optimal position of the current population; dim is the dimension set by the algorithm; λ = 0.3, τ = 2, ξ = 2; t is the current number of iterations, and T is the maximum number of iterations; (5.4) In the algorithm, in order to avoid falling into the local optimum and move towards the optimal solution position, the position update formula of the solution at this stage is: in, is the new position obtained, X Avalanche is the position of other solutions when the local optimal state occurs randomly; (5.5) When a local optimum occurs, the entire algorithm will try its best to save each solution from being trapped. The position update formula is as follows: Among them, X Team is the average position of all solutions; (5.6) The position update formula using the Cauchy mutation operator is: Among them, cauchy(0,1) is the standard Cauchy distribution function; the one-dimensional Cauchy variogram centered at the origin is as follows: Cauchy mutation is used to perturb the individuals in the position update of the climbers and expand the search scale of the climbing team optimization algorithm.

9. The wind turbine generator set small sample fault trend early warning method according to claim 1, characterized in that: The step (6) comprises: (6.1) Assume that there is a data set φ = ((x; t i ), x i ∈R n , t i ∈R m ) N , where x i =[x i1 , x i2 , ..., x in ] T is the network input, t i =[t i1 , t i2 , ..., t in ] T is the network output; the network structure of ELM can be described as: Where G(α i , b i , x) is the output of the i-th hidden layer node; the minimum error is constructed as ||HT||, and by solving the least squares of H=T, we can get β=(H T H) -1 H T T, where: H is the hidden layer node output; T is the label. Based on the RKELM algorithm of prototype clustering preprocessing, the network has more hidden layer nodes; (6.2) Let the training sample be X = [x1, x2, ..., x N1 ], contains N1 samples from C-type targets, and the number of hidden nodes in RKELM is N h , then first perform prototype clustering on X and divide it into N h disjoint training sample clusters, the resulting prototype vector set is Y P =[y1,y2,...,y Nh ],y j is the prototype vector of the jth cluster, j = 1, 2, ..., N h ; The clustering structure of the sample can be represented by the prototype vector set Y P Characterize, and Y P The distance between the two vectors in the P As the support vector of RKELM, the prototype clustering algorithm adopted is K-means++.

10. A wind turbine generator system with a small number of samples fault trend warning system, characterized in that: It includes data acquisition module, data enhancement and sample generation module, fault prediction module, data visualization and fault diagnosis module; The data acquisition module uses a variety of sensors to collect data from various components of the wind turbine in real time, and formulates four fault categories; The data enhancement and sample generation module uses the VAE-DFA-GAN model to generate high-quality original data based on the few sample parameters collected by the data collection module; The fault prediction module uses the VMD-TEMPO-IMTBO model to predict faults on the data of the above modules and generates a 16-step prediction waveform with one point every 15 minutes for 4 hours; The data visualization and fault diagnosis module identifies the fault condition of the wind turbine and the 4-hour 16-step prediction waveform through the extreme learning machine algorithm with prototype clustering preprocessing, and finally displays it in real time through cloud platform data visualization.

Images