A method for training an early icing model of a wind turbine blade and related equipment

Abstract

This application discloses a training method and related equipment for an early icing model of wind turbine blades, relating to the field of wind turbine detection technology. The method includes: acquiring multiple historical SCADA data sets of the wind turbine blades; processing the historical SCADA data to obtain multiple target sample data sets, including preprocessing, dimensionality reduction, and positive / negative sample balancing; and iteratively training an initial model based on the multiple target sample data sets to obtain a target early icing detection model. This solves the problem that early icing samples cannot cover all wind turbine operating conditions and climate environments, leading to insufficient detection accuracy of the early icing detection model.

Description

Technical Field

[0001] This application relates to the field of wind turbine testing technology, and in particular to a method and related equipment for training an early icing model of wind turbine blades. The related equipment includes an early icing model training system for wind turbine blades, a computing device, and a computer-readable storage medium. Background Technology

[0002] Wind turbine blade icing is a complex and dynamic process, influenced by various climatic factors, resulting in diverse icing patterns. Accurate identification of wind turbine icing status is crucial: it provides start-up and shutdown criteria for turbines equipped with de-icing systems, and also provides timely warnings before turbines without de-icing systems start with slight icing, minimizing equipment damage and personnel injury risks. Currently, icing analysis is typically performed by installing sensors at various locations on the wind turbine blades to collect data.

[0003] However, due to the difficulty in installing sensors on the blade itself, the variable environment, equipment limitations, and complex climate, coupled with the strong time-varying nature of the icing process, the collection of early icing data is limited and it is difficult to cover all wind turbine operating conditions and climate environments. This makes accurate observation data for the critical early "slight icing" stage extremely scarce, and the small sample size directly hinders the training and generalization of machine learning and deep learning models, resulting in insufficient detection accuracy of early icing detection models. Summary of the Invention

[0004] To overcome the problem that early icing samples cannot cover all wind turbine operating conditions and climate environments, resulting in insufficient detection accuracy of early icing detection models, this application provides a training method and related equipment for early icing models of wind turbine blades. The related equipment includes an early icing model training system for wind turbine blades, a computing device, and a computer-readable storage medium.

[0005] Firstly, in order to solve the aforementioned technical problems, this application provides a method for training an early icing model for wind turbine blades, comprising: Acquire multiple historical SCADA data of wind turbine blades; Data processing is performed on multiple historical SCADA data to obtain multiple target sample data. The data processing includes preprocessing, data dimensionality reduction, and positive and negative sample balancing. The initial model is iteratively trained based on multiple target sample data to obtain the target early icing detection model.

[0006] Secondly, this application also provides a training system for an early icing model of wind turbine blades, comprising: The acquisition module is used to acquire multiple historical SCADA data of wind turbine blades; The data processing module is used to process data based on multiple historical SCADA data to obtain multiple target sample data. The data processing includes preprocessing, data dimensionality reduction, and positive and negative sample balancing. The model training module is used to iteratively train the initial model based on multiple target sample data to obtain the target early icing detection model.

[0007] Thirdly, this application also provides a computing device, including a memory, a processor, and a program stored in the memory and running on the processor, wherein the processor executes the program to implement the steps of the above-described method for training an early icing model of wind turbine blades.

[0008] Fourthly, this application also provides a computer-readable storage medium storing instructions that, when executed on a terminal device, cause the terminal device to perform the steps of a method for training an early icing model of wind turbine blades.

[0009] The beneficial effects of this application are as follows: First, by performing a series of data processing steps, including preprocessing, dimensionality reduction, and positive / negative sample balancing, based on multiple historical SCADA data of wind turbine blades, high-quality icing-like fault sample data can be derived from a small number of early, minor icing sample data to cover all wind turbine operating conditions and climatic environments, resulting in multiple target sample data. Second, by iteratively training the initial model based on multiple target sample data, gradient vanishing is significantly suppressed and boundary recognition accuracy is improved. This enables the trained early icing detection model to meet the requirements of early icing detection under various wind turbine operating conditions and climatic environments, thereby improving the detection accuracy of the early icing detection model. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating an exemplary embodiment of the present application of a method for training an early icing model of wind turbine blades. Figure 2 This is a schematic flowchart illustrating the early icing model training method for wind turbine blades provided in an exemplary embodiment of this application. Figure 3 This is a schematic diagram of the structure of an early icing model training system for wind turbine blades, as shown in an exemplary embodiment of this application. Figure 4 This is a schematic diagram of the structure of a computer system of a computing device, illustrating an exemplary embodiment of this application. Detailed Implementation

[0011] The following embodiments are further explanations and supplements to this application and do not constitute any limitation on this application.

[0012] In the existing technology, the following early icing methods exist: Nacelle sensors: These sensors directly measure the icing conditions around the wind turbine or outside the nacelle by installing icing sensors. However, due to material limitations and rotor rotation factors, icing sensors cannot provide equivalent feedback on blade icing conditions.

[0013] Blade sensors: Fiber optic sensors are installed on the surface of wind turbine blades to monitor blade strain, and vibration sensors are installed to monitor blade icing through the blade's natural frequency. However, installing sensors increases costs, and because blades weigh tens of tons and have low vibration frequencies, the vibration frequency feedback is not obvious in the initial stage of icing, making it difficult to identify the initial stage of blade icing.

[0014] For turbine operation data: Machine learning algorithms, including supervised and unsupervised methods, are used to identify blade icing conditions. However, due to the complex and highly volatile operating environment of wind turbines and the difficulty in identifying initial icing conditions, accurate identification of the icing status is not possible.

[0015] It is evident that the early icing of existing wind turbine blades is minimal and concealed. Adding sensors increases costs and introduces the risk of drift failure. Traditional machine learning is limited by scarce icing samples, resulting in underfitting models, high false alarm rates, and often triggering de-icing equipment to run idle or the unit to shut down, causing power loss.

[0016] To address the aforementioned issues, embodiments of this application provide a method and related equipment for training an early icing model of wind turbine blades. The related equipment includes an early icing model training system for wind turbine blades, a computing device, and a computer-readable storage medium. These embodiments will be described in detail below.

[0017] The early icing model training method for wind turbine blades provided in this application embodiment can be specifically executed by a server. It should be noted that the server can be an independent server, or it can be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms. No limitation is imposed here.

[0018] This application presents a training method for an early icing model of wind turbine blades. It utilizes a Conditional Generative Adversarial Network (cGAN) to generate high-quality pseudo-icing samples under key operating conditions such as wind speed and power, effectively addressing the challenges of data scarcity and distribution drift (limited data on early minor icing faults and imbalance between healthy and faulty samples). This method can derive high-quality, similar icing fault sample data from a limited pool of early minor icing samples. The generated balanced dataset is then fed into CatBoost to train an efficient algorithm for identifying early minor icing faults in wind turbines. This significantly suppresses gradient vanishing and improves boundary recognition accuracy, providing support for efficient start-stop control strategies for de-icing equipment, ensuring the de-icing system operates only when truly needed. This method is used to address problems with wind turbine blades.

[0019] Please see Figure 1 , Figure 1 An exemplary embodiment of this application illustrates a method for training an early icing model for wind turbine blades, such as... Figure 1 As shown, this application provides a method for training an early icing model for wind turbine blades, including: S11, acquire multiple historical SCADA data of wind turbine blades; S12, based on multiple historical SCADA data, performs data processing to obtain multiple target sample data. The data processing includes preprocessing, data dimensionality reduction processing, and positive and negative sample balancing processing. S13, the initial model is iteratively trained based on multiple target sample data to obtain the target early icing detection model.

[0020] The early icing model training method for wind turbine blades provided in this application firstly involves a series of data processing steps, including preprocessing, dimensionality reduction, and positive / negative sample balancing, based on multiple historical SCADA (Supervisory Control and Data Acquisition) data of the wind turbine blades. This process generates high-quality, similar icing fault sample data from a limited number of early, minor icing samples, covering all wind turbine operating conditions and climate environments, resulting in multiple target sample data sets. Secondly, the initial model is iteratively trained based on these multiple target sample data sets to significantly suppress gradient vanishing and improve boundary recognition accuracy. This ensures that the trained early icing detection model can meet the requirements for early icing detection under various wind turbine operating conditions and climate environments, thereby improving the detection accuracy of the early icing detection model.

[0021] Optionally, data processing is performed based on multiple historical SCADA data to obtain multiple target sample data, including: Multiple historical SCADA data sets were preprocessed to obtain multiple initial sample data sets; For each initial sample data, perform data dimensionality reduction processing to obtain the corresponding dimensionality-reduced sample data; Multiple target sample data are obtained by performing positive and negative sample balancing on multiple dimensionality-reduced sample data.

[0022] In the embodiment provided in this application, firstly, multiple historical SCADA data are preprocessed to remove interference data, resulting in multiple initial sample data. Each initial sample data undergoes dimensionality reduction processing to improve feature quality while reducing data complexity, yielding corresponding dimensionality-reduced sample data. Secondly, positive and negative sample balancing is performed based on the multiple dimensionality-reduced sample data to expand the sample quantity. This ensures that the processed target sample data achieves a balance between positive and negative samples, thereby addressing the problem of scarce and weak features in early icing samples of turbine blades. This facilitates subsequent improvement in the accuracy and generalization ability of model training, enabling the trained early icing detection model to meet the detection accuracy requirements under various wind turbine operating conditions and climatic environments.

[0023] Optionally, multiple historical SCADA data sets are preprocessed to obtain multiple initial sample data sets, including: Based on the preset early icing prerequisites, multiple historical SCADA data are filtered to obtain multiple first sample data. Based on the preset interval deviation conditions, outliers are removed from multiple first sample data to obtain multiple initial sample data.

[0024] In the embodiment provided in this application, multiple historical SCADA data are filtered based on preset early icing prerequisites to initially remove useless data, resulting in multiple first sample data. Then, based on preset interval deviation conditions, outlier removal is performed on these first sample data to further eliminate abnormal data. This improves the accuracy of the obtained initial sample data, thereby increasing the accuracy requirements for subsequent model training and ultimately improving the detection accuracy of the final early icing detection model. Specifically, the early icing prerequisites are an average temperature outside the wind turbine nacelle of less than 2 degrees Celsius for one consecutive minute and an average humidity greater than 80%. The interval deviation conditions are based on temperature, humidity, and wind speed, with interval data statistics calculated at 2°C, 5%, and 0.5 m / s, respectively.

[0025] In this embodiment, historical SCADA data includes continuous feature data, discrete feature data, and basic statistics for each type of feature within a preset time period. The continuous and discrete feature data are shown in Table 1, and the specific data range for each data point can be adjusted according to the actual wind turbine. Specifically, the continuous feature data includes: wind speed, active power, blade 1 angle, blade 2 angle, blade 3 angle, blade 1 motor current, blade 2 motor current, blade 3 motor current, nacelle outside temperature, nacelle outside humidity, speed, and the statistical values ​​of the 10-minute maximum, minimum, average, and standard deviation of torque. The discrete feature data includes digital grid-connected / disconnected, fault / non-fault, power curtailment / non-power curtailment, seasonal variables, and whether icing occurs. The basic statistics include mean, maximum, minimum, and standard deviation statistics, with a preset time period of 10 minutes. In an exemplary embodiment provided in this application, multiple historical SCADA data are preprocessed to obtain multiple initial sample data. The specific steps are as follows: Two main categories of second-level SCADA data are acquired: continuous feature data and discrete feature data of wind turbine blades. Using a 1-minute step sliding window, the mean, maximum, minimum, and standard deviation of each feature in the historical SCADA data are calculated within 10 minutes to form basic statistics. Based on the continuous feature data, discrete feature data, and the basic statistics of each feature within a preset time period, historical SCADA data is generated and used as input for early slight icing discrimination of wind turbine blades.

[0026] To improve model operating efficiency, reduce model training costs, and enhance model training accuracy, temperature and humidity were used as start-up discriminant variables and data input criteria. The average temperature outside the wind turbine nacelle for one consecutive minute being less than 2 degrees Celsius and the average humidity being greater than 80% were used as prerequisites for early icing and as input conditions for model training data. Multiple historical SCADA data were filtered to obtain multiple first sample data.

[0027] Outlier data removal: Since temperature, humidity, and wind speed are the most important variables affecting icing and fan operation, during the training phase, based on the preset interval deviation conditions, multiple first sample data are statistically analyzed according to temperature, humidity, and wind speed at intervals of 2℃, 5%, and 0.5m / s, respectively. If any data exceeds 10 times the standard deviation, it is considered outlier and is deleted, resulting in multiple initial sample data.

[0028] Optionally, dimensionality reduction processing is performed on each initial sample data to obtain the corresponding dimensionality-reduced sample data, including: Input multiple initial sample data into a preset variational Gaussian mixture model, and calculate the probability density of each initial sample data under each single mode distribution; For each pattern in each probability density, perform pattern normalization to obtain the corresponding pattern vector; Multiple pattern vectors from each initial sample data are concatenated to obtain the corresponding dimensionality-reduced sample data.

[0029] In the embodiment provided in this application, multiple initial sample data are input into a preset variational Gaussian mixture model. First, the probability density of each single-mode distribution of each initial sample data is calculated, and each mode in each probability density is normalized to obtain the corresponding mode vector. Then, multiple mode vectors based on each initial sample data are concatenated to obtain the corresponding dimensionality-reduced sample data. In this way, the variational Gaussian mixture model (also known as the non-variable Gaussian mixture model) is used to solve the problems of data dimensionality reduction and multi-modal distribution, so as to improve the feature quality of the data while reducing data complexity. This can improve training efficiency and training targeting in subsequent model training, and thus improve the detection accuracy of the final target early icing detection model. Here, the mode vector is represented by the one-hot vector to which the mode belongs and the mode normalization value. The number of mode vectors mn is determined by the number of single-mode distributions m and the number of modes n.

[0030] In an exemplary embodiment provided in this application, to address the high dimensionality and multi-peak distribution characteristics of SCADA data, a variational Gaussian mixture model is used for clustering dimensionality reduction and feature reshaping, which can simplify the input space of the subsequent neural network to a certain extent. Specifically, the variational Gaussian mixture model is used to perform dimensionality reduction processing on each initial sample data to obtain the corresponding dimensionality-reduced sample data. The specific steps are as follows: First, calculate the probability density of each single-pattern distribution for each initial sample data: ① For each continuous feature data in multiple initial sample data The number of individual mode distributions is estimated using a variational Gaussian mixture model. The estimation formula is as follows: in, It is a Gaussian mixture model; For the first The number of patterns for a continuous feature; and These are the first and second parts of the single-mode distribution. Weights and standard deviations (variance) of each pattern; For the first in a single-pattern distribution The mean of each pattern.

[0031] ②For Each value in Calculate its probability density under each single-mode distribution. The corresponding calculation formula is as follows: Then, for each pattern in each probability density, pattern normalization is performed to obtain the corresponding pattern vector: a pattern is extracted from the given probability density and normalized, and the features... Each value in Both are represented by the one-hot vector to which the mode belongs and the mode normalization value, that is, using and These represent the normalized data value within the pattern distribution and the one-hot encoded data index of the selected pattern distribution, respectively. For example, when the number of patterns in a single pattern distribution... When the variational Gaussian mixture model has three modes, if the first mode is selected, then... and The calculation formula is: in, As a one-hot vector, It is a one-hot encoding mode.

[0032] Finally, the multiple pattern vectors from each initial sample data are concatenated to obtain the corresponding dimensionality-reduced sample data. Specifically, it can be expressed as: in This is the normalized value of the j-th data point of the first feature; One-hot encoding of the j-th data of the first numerical feature; The normalized value of the j-th data of the Nc-th feature; One-hot encoding of the j-th data point of the first digital feature; is the one-hot encoding of the j-th data of the Nd-th feature; ⊕ is the concatenation operation in mathematical definition.

[0033] Optionally, positive and negative sample equalization is performed based on multiple dimensionality-reduced sample data to obtain multiple target sample data, including: Multiple dimensionality-reduced sample data are input into the initial adversarial network for positive and negative sample equalization processing to obtain multiple transition sample data. Based on the preset first loss function, the initial adversarial network is iteratively optimized using multiple transitional sample data to obtain the target adversarial network; The target adversarial network outputs the target sample data corresponding to each dimensionality-reduced sample data.

[0034] In the embodiment provided in this application, multiple dimensionality-reduced sample data are input into an initial adversarial network for positive and negative sample balancing to obtain multiple transitional sample data. Based on a preset first loss function, the initial adversarial network is iteratively optimized using the multiple transitional sample data to obtain a target adversarial network and multiple target sample data output to produce pseudo early icing sample data that meets accuracy requirements. This expands the number of samples so that the multiple target sample data can achieve positive and negative sample balancing, thereby solving the problem of scarce early icing samples and weak features of generator blades. This facilitates subsequent improvement of the accuracy and generalization ability of model training, and enables the trained target early icing detection model to meet the detection accuracy requirements of various wind turbine operating conditions and climatic environments.

[0035] In this embodiment, the expression for the first loss function is: ,in, The expected value of the real data follows an x~pdata distribution; Let G be the expected value of the noise data, and let the noise follow a distribution x ~ pz; G is the generator, which produces pseudo sample data based on the distribution of z; D is the discriminator, which judges the generated fake data based on the data learning pattern of the real data x distribution; D(G(z)) is the result of the discriminator judging the fake data generated by the generator.

[0036] Optionally, the initial adversarial network includes a generative model and a discriminative model; multiple dimensionality-reduced sample data are input into the initial adversarial network for positive and negative sample equalization to obtain multiple transitional sample data, including: Multiple dimensionality-reduced sample data are input into the initial adversarial network, and noise introduction and condition processing are performed on each dimensionality-reduced sample data to obtain enhanced sample data. Each augmented sample data is processed using a generative model to obtain the corresponding generated data; Each dimensionality-reduced sample data is subjected to pattern normalization to obtain the true data. The probability of true or false data is obtained by using a discriminant model to calculate the difference between generated data and corresponding real data. Multiple transitional sample data are formed based on multiple generated data and multiple dimensionality-reduced sample data corresponding to the true and false probabilities that meet the preset probability requirements.

[0037] In the embodiment provided in this application, multiple dimensionality-reduced sample data are input into the initial adversarial network. First, noise introduction processing and condition processing are performed on each dimensionality-reduced sample data. Noise introduction processing can achieve data augmentation (artificially expanding the dataset to prevent overfitting), robust training (allowing the model to learn to ignore interference and focus on essential features), and simulation of reality (making the data closer to the imperfect real world). Condition processing controls the generation / transformation: guiding data transformation or generation under specific conditions to achieve data augmentation and obtain augmented sample data. Secondly, the generative model is used to process each enhanced sample data to produce pseudo early icing sample data that meets the accuracy requirements, thus obtaining the corresponding generated data. The discriminative model is then used to calculate the difference between the generated data and the corresponding real data to obtain the true and false probabilities. Based on the true and false probabilities that meet the preset probability requirements, multiple generated data and multiple dimensionality-reduced sample data are formed to form multiple transitional sample data, thereby achieving precise expansion of the number of samples. This ensures that the positive and negative samples of multiple target sample data can be balanced, solving the problem of scarce early icing samples and weak features of turbine blades. This facilitates the subsequent improvement of the accuracy and generalization ability of model training, enabling the trained target early icing detection model to meet the detection accuracy requirements of various wind turbine operating conditions and climatic environments.

[0038] In an exemplary embodiment provided in this application, in order to address the problem of scarce early icing samples and weak features of wind turbine blades, a Generative Adversarial Network (GAN) is used to generate a number of pseudo-icing samples equal to the number of real fault samples, thereby achieving a balance between positive and negative samples and improving the accuracy and generalization ability of the subsequent classification model.

[0039] Specifically, generative adversarial networks (GANs) originate from the concept of "zero-sum games" in game theory, and mainly consist of two models: a generative model and a discriminative model. During training, the goal of the generative model is to make the distribution of the generated data approximate the distribution of the real data; while the goal of the discriminative model is to identify as many differences as possible between the data generated by the generative model and the distribution of the real data.

[0040] The specific network setup for adversarial networks: ① Network structure of generative model G: Noise 128 dimensional ⊕ Condition 8 dimensional ⊕ 136 dimensional ⊕ Projection 512 dimensional ⊕ ResBlock (residual block) × 4 ⊕ Split head. Wherein, noise z∈R128, and condition c∈R4 (4, one-hot).

[0041] Continuous head: Outputs 96 dimensions, using the tanh activation function (-1, 1).

[0042] Discrete head: 5 softmax groups (5 groups of 2 classes) (discrete features).

[0043] Output: Continuous part x_c∈R96, discrete part π=[π1,…,π4] (4 sets of log-softmax vectors).

[0044] As can be seen, G: 1 projection + 4 ResBlock × 2 layers + 2 output heads = 11 layers.

[0045] ② Network structure of discriminant model D: 108-dimensional sample ⊕ 8-dimensional condition → 116-dimensional → 512 → ResBlock × 4 → single-node sigmoid.

[0046] Output: scalar s∈(0,1) represents the "true" probability.

[0047] As can be seen, D: 1 input + 4 ResBlocks × 2 layers + 1 output = 9 layers. In addition, all fully connected layers in the generative model G and the discriminative model D are not biased by default.

[0048] Optionally, the initial model is a CatBoost model; the initial model is iteratively trained based on multiple target sample data to obtain an early icing detection model for the target, including: Multiple target sample data are input into the CatBoost model to predict early icing and obtain early icing results. Using a pre-defined second loss function, the CatBoost model is iteratively optimized using multiple early icing results to obtain the target early icing detection model.

[0049] In the embodiment provided in this application, the CatBoost model is trained using multiple target sample data containing early slight icing under various wind turbine operating conditions and climate environments, as well as a preset second loss function. This can significantly suppress gradient vanishing and improve boundary recognition accuracy, so that the trained early icing detection model can meet the requirements of early icing detection under various wind turbine operating conditions and climate environments, thereby improving the detection accuracy of the early icing detection model.

[0050] In this embodiment, the expression for the second loss function is: , where y is the categorical data label; p is the probability that the model predicts a positive value.

[0051] In one exemplary embodiment provided in this application, the CatBoost model incorporates the CatBoost algorithm for classification to verify real-world icing data. This CatBoost algorithm is an improvement upon the GBDT algorithm, and its accuracy surpasses that of the XGBoost and LightGBM algorithms, which belong to the same GBDT framework. The CatBoost algorithm replaces the gradient estimation method in traditional algorithms with a ranking boosting method, solving the problems of gradient bias and prediction offset. It also uses an Oblivious tree (symmetric tree) as the base model, improving the model's correct classification ability while also considering generalization ability, effectively preventing overfitting.

[0052] Specifically, the CatBoost model is used as follows: ① Support for categorical variables: A method based on the statistical value of the prediction target is designed using CatBoost. All data samples are randomly permuted to generate multiple random sequences. For a given feature sequence, the average label value is used to replace the category during training. Assume... If the dataset is a reordered sequence, then the sequence from the original dataset can be... The first sample Features use express: Where σ represents the newly created sequential dataset; P represents the position of the current sample in the permutation, or the Pth value; k represents the kth feature; x σp,k Y is the k-th feature of the p-th sample; σj Let be the target value of the j-th sample; p be the prior value; and a be the smoothing parameter.

[0053] This method can transform categorical features into numerical features, reducing computation and minimizing information loss.

[0054] ② Sorting boosting method: The CatBoost algorithm proposes the idea of ​​sorting boosting, which can effectively reduce gradient estimation error and alleviate prediction offset problem.

[0055] ③ Oblivious Tree (Symmetric Tree): CatBoost uses a symmetric binary tree as its base model. This tree structure constraint has a certain regularization effect. In the CatBoost prediction process, the splitting of each feature is independent and there is no order in which they are split. Multiple samples can be predicted simultaneously, improving the prediction speed of the CatBoost model.

[0056] ④ Second loss function: Uses the binary classification cross-entropy loss function It can be fine-tuned with sample weights or focal loss, and is particularly effective for fault detection with small samples.

[0057] Please see Figure 2 , Figure 2 This is a schematic flowchart illustrating the application of the early icing model training method for wind turbine blades provided in an exemplary embodiment of this application, as shown below. Figure 2 As shown, the specific application process of the early icing model training method for wind turbine blades is as follows: Data preprocessing: Multiple historical SCADA data sets are preprocessed to obtain multiple initial sample data sets; Sliding window hybrid feature construction: Perform dimensionality reduction on each initial sample data to obtain the corresponding dimensionality-reduced sample data; Positive and negative sample balancing: Noise introduction and condition processing are applied to each dimensionality-reduced sample data to obtain enhanced sample data; a generative model is used to process each enhanced sample data to obtain corresponding generated data; pattern normalization is performed on each dimensionality-reduced sample data to obtain real data; a discriminative model is used to calculate the difference between the generated data and the corresponding real data to obtain the true / false probability; multiple transitional sample data are formed based on multiple generated data and multiple dimensionality-reduced sample data corresponding to the true / false probabilities that meet the preset probability requirements. Model training: Input multiple target sample data into the CatBoost model, and use the preset second loss function to iteratively optimize the CatBoost model to obtain a target early icing detection model that meets the detection accuracy requirements. Then, use the target early icing detection model to perform early icing diagnosis and detection of wind turbine blades.

[0058] Please see Figure 3 , Figure 3 An exemplary embodiment of this application illustrates a training system for an early icing model of wind turbine blades, such as... Figure 3 As shown, this application provides a training system 300 for an early icing model of wind turbine blades, comprising: The acquisition module 301 is used to acquire multiple historical SCADA data of the wind turbine blades; Data processing module 302 is used to process data based on multiple historical SCADA data to obtain multiple target sample data. The data processing includes preprocessing, data dimensionality reduction processing, and positive and negative sample balancing processing. The model training module 303 is used to iteratively train the initial model based on multiple target sample data to obtain the target early icing detection model.

[0059] The early icing model training system 300 for wind turbine blades provided in this application firstly utilizes a data processing module 302 to perform a series of data processing steps based on multiple historical SCADA data of the wind turbine blades acquired by the acquisition module 301. These steps include preprocessing, data dimensionality reduction, and positive / negative sample balancing. This allows for the generation of high-quality, similar icing fault sample data from a limited number of early, minor icing sample data, covering all wind turbine operating conditions and climatic environments, resulting in multiple target sample data. Secondly, the model training module 303 iteratively trains the initial model based on the multiple target sample data to significantly suppress gradient vanishing and improve boundary recognition accuracy. This ensures that the trained early icing detection model can meet the requirements for early icing detection under various wind turbine operating conditions and climatic environments, thereby improving the detection accuracy of the early icing detection model.

[0060] Optionally, the data processing module 302 is specifically used for: Multiple historical SCADA data sets were preprocessed to obtain multiple initial sample data sets; For each initial sample data, perform data dimensionality reduction processing to obtain the corresponding dimensionality-reduced sample data; Multiple target sample data are obtained by performing positive and negative sample balancing on multiple dimensionality-reduced sample data.

[0061] Optionally, the data processing module 302 is specifically used for: Based on the preset early icing prerequisites, multiple historical SCADA data are filtered to obtain multiple first sample data. Based on the preset interval deviation conditions, outliers are removed from multiple first sample data to obtain multiple initial sample data.

[0062] Optionally, the data processing module 302 is specifically used for: Input multiple initial sample data into a preset variational Gaussian mixture model, and calculate the probability density of each initial sample data under each single mode distribution; For each pattern in each probability density, perform pattern normalization to obtain the corresponding pattern vector; Multiple pattern vectors from each initial sample data are concatenated to obtain the corresponding dimensionality-reduced sample data.

[0063] Optionally, the data processing module 302 is specifically used for: Multiple dimensionality-reduced sample data are input into the initial adversarial network for positive and negative sample equalization processing to obtain multiple transition sample data. Based on the preset first loss function, the initial adversarial network is iteratively optimized using multiple transitional sample data to obtain the target adversarial network; The target adversarial network outputs the target sample data corresponding to each dimensionality-reduced sample data.

[0064] Optionally, the initial adversarial network includes a generative model and a discriminative model; the data processing module 302 is specifically used for: Multiple dimensionality-reduced sample data are input into the initial adversarial network, and noise introduction and condition processing are performed on each dimensionality-reduced sample data to obtain enhanced sample data. Each augmented sample data is processed using a generative model to obtain the corresponding generated data; Each dimensionality-reduced sample data is subjected to pattern normalization to obtain the true data. The probability of true or false data is obtained by using a discriminant model to calculate the difference between generated data and corresponding real data. Multiple transitional sample data are formed based on multiple generated data and multiple dimensionality-reduced sample data corresponding to the true and false probabilities that meet the preset probability requirements.

[0065] Optionally, the initial model is a CatBoost model; model training module 303 is specifically used for: Multiple target sample data are input into the CatBoost model to predict early icing and obtain early icing results. Using a pre-defined second loss function, the CatBoost model is iteratively optimized using multiple early icing results to obtain the target early icing detection model.

[0066] It should be noted that the early icing model training system for wind turbine blades provided in the above embodiments and the early icing model training method for wind turbine blades provided in the above embodiments belong to the same concept. The specific operation methods of each module and unit have been described in detail in the method embodiments and will not be repeated here. In practical applications, the early icing model training system for wind turbine blades provided in the above embodiments can be assigned to different functional modules as needed, that is, the internal structure of the system can be divided into different functional modules to complete all or part of the functions described above. This is not a limitation here.

[0067] A computing device according to an embodiment of this application includes a memory, a processor, and a program stored in the memory and running on the processor. When the processor executes the program, it implements some or all of the steps of the above-described method for training an early icing model of wind turbine blades.

[0068] The computing device can be a computer, and the corresponding program is computer software. The parameters and steps in the computing device described above can be referred to the parameters and steps in the embodiment of the early icing model training method for wind turbine blades mentioned above, and will not be repeated here.

[0069] Figure 4 A schematic diagram of a computer system suitable for implementing the embodiments of this application is shown. It should be noted that... Figure 4 The computer system 400 shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0070] like Figure 4 As shown, the computer system 400 includes a Central Processing Unit (CPU) 401, which can perform various appropriate actions and processes, such as executing the methods described in the above embodiments, based on programs stored in Read-Only Memory (ROM) 402 or programs loaded from Storage Unit 408 into Random Access Memory (RAM) 403. The RAM 403 also stores various programs and data required for system operation. The CPU 401, ROM 402, and RAM 403 are interconnected via a bus 404. An Input / Output (I / O) interface 405 is also connected to the bus 404.

[0071] The following components are connected to I / O interface 405: an input section 406 including a keyboard, mouse, etc.; an output section 407 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 408 including a hard disk, etc.; and a communication section 409 including a network interface card such as a LAN (Local Area Network) card, modem, etc. The communication section 409 performs communication processing via a network such as the Internet. A drive 410 is also connected to I / O interface 405 as needed. A removable medium 411, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 410 as needed so that computer programs read from it can be installed into storage section 408 as needed.

[0072] Specifically, according to embodiments of this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program including a computer program for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 409, and / or installed from removable medium 411. When the computer program is executed by central processing unit (CPU) 401, it performs various functions defined in the system of this application.

[0073] This application embodiment provides a computer-readable storage medium storing instructions that, when executed, perform the steps of the aforementioned method for training an early icing model for wind turbine blades. The computer-readable storage medium can be either transient or non-transient.

[0074] The technical solutions of this disclosure can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes one or more instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of this disclosure. The aforementioned computer-readable storage medium can be a non-transitory computer-readable storage medium, including: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, and other media capable of storing program code; it can also be a transient computer-readable storage medium.

[0075] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. Each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0076] Those skilled in the art will recognize that this application can be implemented as a system, method, or computer program product. Therefore, this disclosure can be implemented in the following forms: it can be entirely hardware, entirely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software, generally referred to herein as a "module" or "system." Furthermore, in some embodiments, this application can also be implemented as a computer program product contained in one or more computer-readable media, which contains computer-readable program code. Computer-readable storage media can be, for example, but not limited to—electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any combination thereof.

[0077] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0078] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A training method for an early icing model of wind turbine blades, characterized in that, include: Acquire multiple historical SCADA data of wind turbine blades; Data processing is performed on multiple historical SCADA data to obtain multiple target sample data. The data processing includes preprocessing, data dimensionality reduction processing, and positive and negative sample balancing processing. The initial model is iteratively trained based on multiple target sample data to obtain the target early icing detection model.

2. The method according to claim 1, characterized in that, The data processing based on multiple historical SCADA data points yields multiple target sample data, including: Multiple historical SCADA data are preprocessed to obtain multiple initial sample data; For each initial sample data, perform data dimensionality reduction processing to obtain the corresponding dimensionality-reduced sample data; Based on the multiple dimensionality-reduced sample data, positive and negative sample balancing is performed to obtain multiple target sample data.

3. The method according to claim 2, characterized in that, The process of preprocessing multiple historical SCADA data sets yields multiple initial sample data sets, including: Based on the preset early icing prerequisite, multiple historical SCADA data are filtered to obtain multiple first sample data. Based on the preset interval deviation condition, outliers are removed from multiple first sample data to obtain multiple initial sample data.

4. The method according to claim 2, characterized in that, The step of performing dimensionality reduction processing on each initial sample data to obtain the corresponding dimensionality-reduced sample data includes: Input multiple initial sample data into a preset variational Gaussian mixture model, and calculate the probability density of each initial sample data under each single mode distribution; For each pattern in each probability density, perform pattern normalization to obtain the corresponding pattern vector; Multiple pattern vectors from each initial sample data are concatenated to obtain the corresponding dimensionality-reduced sample data.

5. The method according to claim 2, characterized in that, The positive and negative sample balancing process based on multiple dimensionality-reduced sample data results in multiple target sample data, including: Multiple dimensionality-reduced sample data are input into the initial adversarial network for positive and negative sample equalization processing to obtain multiple transition sample data. Based on a preset first loss function, the initial adversarial network is iteratively optimized using multiple transitional sample data to obtain the target adversarial network; The target adversarial network outputs the target sample data corresponding to each dimensionality-reduced sample data.

6. The method according to claim 5, characterized in that, The initial adversarial network includes a generative model and a discriminative model; the input of multiple dimensionality-reduced sample data into the initial adversarial network for positive and negative sample equalization processing yields multiple transitional sample data, including: Multiple dimensionality-reduced sample data are input into the initial adversarial network, and noise introduction and condition processing are performed on each dimensionality-reduced sample data to obtain enhanced sample data. The generative model is used to process each augmented sample data to obtain the corresponding generated data; Each dimensionality-reduced sample data is subjected to pattern normalization to obtain the true data. The discriminant model is used to calculate the difference between the generated data and the corresponding real data to obtain the probability of true or false data. Multiple transitional sample data are formed based on multiple generated data corresponding to true and false probabilities that meet preset probability requirements and multiple dimensionality-reduced sample data.

7. The method according to any one of claims 1 to 6, characterized in that, The initial model is a CatBoost model; the iterative training of the initial model based on multiple target sample data to obtain the early icing detection model includes: Multiple target sample data are input into the CatBoost model to predict early icing and obtain early icing results. Using a preset second loss function, the CatBoost model is iteratively optimized using multiple early icing results to obtain the target early icing detection model.

8. A training system for an early icing model of wind turbine blades, characterized in that, include: The acquisition module is used to acquire multiple historical SCADA data of wind turbine blades; The data processing module is used to process data based on multiple historical SCADA data to obtain multiple target sample data. The data processing includes preprocessing, data dimensionality reduction processing, and positive and negative sample balancing processing. The model training module is used to iteratively train the initial model based on multiple target sample data to obtain the target early icing detection model.

9. A computing device comprising a memory, a processor, and a program stored in the memory and running on the processor, characterized in that, When the processor executes the program, it implements the steps of the early icing model training method for wind turbine blades as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed on a terminal device, cause the terminal device to perform the steps of the early icing model training method for wind turbine blades as described in any one of claims 1 to 7.

Images