A robust noise reduction method for wind turbine data based on information theory learning

Abstract

This invention discloses a robust denoising method for wind turbine data based on information theory learning. The method includes: acquiring historical wind turbine data under different operating conditions; constructing an autoencoder model containing a cascaded encoder and decoder; constructing a denoising loss function and a feature regularization loss function; directly using the acquired historical wind turbine data as training data, weighting the denoising loss function and the feature regularization loss function to construct the total loss function of the autoencoder model, and iteratively training the autoencoder model using the training data itself as the reconstruction target; inputting the wind turbine data to be denoised into the trained autoencoder model for processing to obtain the denoised wind turbine data. This invention's method does not require the use of clean, noise-free data during training, and the model training process exhibits good noise robustness, enabling effective denoising of wind turbine data.

Description

Technical Field

[0001] This invention belongs to the field of interdisciplinary technology of machine learning and new energy, and specifically relates to a robust noise reduction method for wind turbine data based on information theory learning. Background Technology

[0002] During wind turbine operation, data such as speed, current, and voltage generated are crucial for wind turbine condition monitoring, fault diagnosis, and operational optimization. However, the complex operating environment of wind turbines makes the collected data susceptible to electromagnetic interference and environmental disturbances, introducing significant noise. This noise degrades data quality, impacting the accuracy of subsequent processing such as condition monitoring. Therefore, effective noise reduction processing of wind turbine data is necessary.

[0003] In recent years, autoencoders have been widely used in data denoising tasks. Autoencoder-based denoising methods compress input data into low-dimensional latent features through an encoder, and then map these low-dimensional latent features back to the input dimension through a decoder. Based on the information bottleneck principle, autoencoders can preserve the main structural information of the data while suppressing random and unstable components, thus achieving noise suppression. Traditional autoencoder-based data denoising methods typically employ a denoising autoencoder architecture, which involves artificially adding noise to clean data and training a model to recover clean data. These methods usually require clean data without noise during training and use mean squared error as the loss function, assuming the noise follows a Gaussian distribution. However, during wind turbine operation, the collected data often inevitably contains noise, and may even contain impulse interference, resulting in a lack of clean data during model training. Furthermore, traditional methods typically use mean squared error as the loss function, assuming the noise approximates a Gaussian distribution. When the actual noise exhibits non-Gaussian characteristics, such as containing impulse interference, the mean squared error becomes overly sensitive to this noise, and the model's robustness is significantly insufficient. Summary of the Invention

[0004] To address the problems existing in the background technology, this invention provides a robust denoising method for wind turbine data based on information theory learning. This method solves the technical problems of poor denoising performance and insufficient robustness in existing wind turbine data denoising methods when clean training data is lacking and noise exhibits non-Gaussian characteristics. The method of this invention is a data-driven signal denoising technique. By constructing a denoising loss function based on maximizing correlation entropy, the model can be trained using only noisy data and effectively suppresses the interference of non-Gaussian noise. Simultaneously, by constructing a feature regularization loss function based on maximizing Renyi entropy, it guides the uniform distribution of latent features, improving the model's ability to preserve data details and its generalization performance.

[0005] The technical solution adopted in this invention is:

[0006] I. A robust noise reduction method for wind turbine data based on information theory learning:

[0007] S1. Obtain wind turbine data under historical wind turbine operating conditions;

[0008] S2. Construct an autoencoder model that includes a cascaded encoder and decoder;

[0009] S3. Construct a noise reduction loss function based on maximizing the relevance entropy and a feature regularization loss function based on maximizing the Renyi entropy, respectively.

[0010] A noise reduction loss function based on maximizing relevance entropy is used to assign adaptive weights to the reconstruction error of the autoencoder model during training, in order to suppress the impact of non-Gaussian noise on model training.

[0011] The Renyi entropy-maximizing feature regularization loss function is used to guide the latent features of the autoencoder model to be uniformly distributed in the feature space during training.

[0012] S4. No need to label the wind turbine data. The historical wind turbine data obtained directly is used as training data. The noise reduction loss function and the feature regularization loss function are weighted and combined to construct the total loss function of the autoencoder model. The training data itself is used as the reconstruction target to iteratively train the autoencoder model and obtain the trained autoencoder model.

[0013] S5. Input the wind turbine data to be denoised into the trained autoencoder model for processing to obtain the denoised wind turbine data.

[0014] The noisy fan data is distorted due to factors such as electromagnetic interference, environmental interference, and pulse interference. The noise-reduced fan data is used for subsequent analysis tasks such as condition monitoring, fault diagnosis, and operation optimization.

[0015] Step S1 specifically involves:

[0016] Multiple operating parameters of historical wind turbine operation status are collected. The operating parameters collected at the same time are combined to construct a single sample of wind turbine data. Wind turbine data samples from consecutive time points constitute wind turbine data. The obtained wind turbine data is standardized to obtain standardized wind turbine data, which is used as training data for the model.

[0017] The operating parameters include wind speed, wind direction, ambient temperature, unit operating temperature, and power.

[0018] Both the encoder and decoder are composed of multiple fully connected layers connected in sequence. The encoder receives the input fan data, maps it through the multi-layer fully connected layers of the encoder, and outputs latent features. The decoder receives the latent features and outputs noise-reduced fan data.

[0019] Each fully connected layer contains multiple neurons, and each fully connected layer corresponds to a set of trainable weight parameters and bias parameters. Each fully connected layer is equipped with a non-linear activation function. The non-linear activation function includes one or more of ReLU, LeakyReLU, Sigmoid, or Tanh.

[0020] The noise reduction loss function based on maximizing the correlation entropy is set according to the following formula:

[0021] ;

[0022] in, Here is the noise reduction loss function; For indexing; The number of training samples; and The first Each training sample and its corresponding model output; For the first The weights corresponding to each training sample; This is the preset kernel function bandwidth parameter; The square of the Euclidean distance between the vectors; It is an exponential function.

[0023] The feature regularization loss function based on maximizing Renyi entropy is set according to the following formula:

[0024]

[0025]

[0026] in, The feature regularization loss function; It is a logarithmic function; For indexing; The first The weights corresponding to each training sample; and The first The and the first The latent features output by the encoder after processing each training sample; It is the square of the Euclidean distance between the vectors.

[0027] When iteratively training the autoencoder model, only noisy wind turbine data is used as training data, and the training data itself is used as the reconstruction target. There is no need to use clean wind turbine data without noise, nor is there any labeled data.

[0028] II. A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method.

[0029] 3. A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above method.

[0030] The innovation of this invention lies not only in proposing a noise reduction loss function and a feature regularization loss function, but also in the fact that the method is an unsupervised approach. This approach addresses the engineering bottleneck of obtaining clean data in real-world wind turbine scenarios, thereby improving the practicality and applicability of the method. Furthermore, unlike supervised learning methods, this invention achieves good noise reduction results without requiring clean data at all.

[0031] The beneficial effects of this invention are:

[0032] (1) The method of the present invention is an unsupervised learning method, which can complete the model training without the need for clean wind turbine data without noise, overcomes the engineering bottleneck that it is difficult to obtain clean data in actual wind turbine scenarios, and improves the practicality and applicability of the method.

[0033] (2) This invention effectively suppresses the negative impact of non-Gaussian noise (such as impulse interference) on model training by using a noise reduction loss function based on maximizing the relevant entropy. Compared with the traditional method based on mean square error as the loss function, it has stronger robustness.

[0034] (3) This invention guides the uniform distribution of potential features by using a feature regularization loss function based on Renyi entropy maximization, thereby improving the model's ability to retain data details and its generalization performance, and thus improving the noise reduction performance.

[0035] (4) The verification of the embodiments shows that the noise reduction performance of the method of the present invention is significantly better than the traditional method that uses mean square error as the loss function under the condition of mixed Gaussian noise and impulse noise, which provides strong support for the reliable acquisition and subsequent analysis of wind turbine data. Attached Figure Description

[0036] Figure 1 This is a flowchart of the method of the present invention.

[0037] Figure 2 This is a diagram showing the composition of the total loss function in this invention.

[0038] Figure 3 This is a structural diagram of the autoencoder model of the present invention.

[0039] Figure 4 This is a comparison chart of the noise reduction results of the method of the present invention and the traditional method. Detailed Implementation

[0040] The present invention will now be described in more detail with reference to the accompanying drawings and embodiments. However, the present invention is not limited thereto. For those skilled in the art, several improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention. Contents not described in detail in this specification are prior art known to those skilled in the art.

[0041] Example 1:

[0042] like Figure 1 As shown, the robust noise reduction method for wind turbine data in this embodiment is implemented according to the following steps:

[0043] S1. Obtain wind turbine data under real historical operating conditions and represent each wind turbine data sample as a multi-dimensional feature vector.

[0044] The system collects real-time historical operating parameters of wind turbines, including wind speed, wind direction, ambient temperature, turbine operating temperature, and power. Operating parameters collected at the same time are combined to form individual samples of wind turbine data, each sample represented as a multi-dimensional feature vector.

[0045]

[0046] in, This is a single sample of wind turbine data; , and They are the 1st, 2nd, and 3rd in a single sample, respectively. The values ​​corresponding to each running parameter Represents the vector dimension.

[0047] Wind turbine data samples from consecutive time points constitute wind turbine data. The obtained wind turbine data is standardized to obtain standardized wind turbine data, which is then used as training data for the model.

[0048] like Figure 3 As shown in Figure S2, a noise-robust autoencoder model for wind turbine data noise reduction is established.

[0049] The noise-robust autoencoder model in this embodiment includes encoders connected in series. and decoder Both the encoder and decoder are composed of multiple fully connected layers connected sequentially. The encoder receives the input fan data, maps it through the encoder's multiple fully connected layers, and outputs latent features. The decoder receives the latent features and outputs noise-reduced fan data.

[0050] In practice, the number of fully connected layers in the encoder and decoder can be the same or different.

[0051] Let a single sample be the input model. encoder Used to perform multi-layer parameter mapping on model inputs and output latent features. ,in decoder latent features Mapping to model output .

[0052] The encoder consists of several fully connected layers connected in series. Each fully connected layer contains multiple neurons, and each layer corresponds to a set of trainable weight and bias parameters. Each fully connected layer also has a non-linear activation function. The overall parameter set of the encoder is denoted as: ,in and These represent the first and second parts of the encoder, respectively. The weight and bias parameters corresponding to each fully connected layer This indicates the number of fully connected layers in the encoder.

[0053] The decoder also consists of several sequentially connected fully connected layers. Each fully connected layer is configured with corresponding weight parameters and bias parameters, and is equipped with a non-linear activation function. The overall parameter set of the decoder is denoted as: ,in and These represent the first and second parts of the decoder, respectively. The weight and bias parameters corresponding to each fully connected layer This indicates the number of fully connected layers in the decoder.

[0054] Activation functions include, but are not limited to, ReLU, LeakyReLU, Sigmoid, and Tanh. Given the difficulty in obtaining noise-free, clean wind turbine data in real-world scenarios, autoencoder models can perform wind turbine data denoising without requiring training on clean wind turbine data.

[0055] S3. Construct a noise reduction loss function based on information theory learning for wind turbine data noise reduction.

[0056] To address the issue that wind turbine data may contain Gaussian and non-Gaussian noise during training, this invention, within an information-theoretic learning framework, uses correlation entropy as an information-theoretic metric to construct a noise reduction loss based on maximizing correlation entropy, thereby enhancing the model's robustness to noisy data. The correlation entropy of the reconstruction error is weighted using a kernel function, ensuring the model optimization process remains highly sensitive to samples with smaller reconstruction errors, while automatically reducing the weight of samples with larger reconstruction errors. This mitigates the impact of noise on the model's learning process and enhances its robustness to noisy data.

[0057] Let the sample set used to train the model and the corresponding model output set be respectively. and ,in and They represent the first Each input sample and its corresponding model output This represents the number of training samples. According to the definition of relevance entropy in information-theoretic learning, for... and The correlation entropy of the reconstruction error is:

[0058]

[0059] in, To reconstruct the correlation entropy of the error; For indexing; It is an exponential function; and The first Each training sample and its corresponding model output; This is the preset kernel function bandwidth parameter; It is the square of the Euclidean distance between the vectors.

[0060] Based on the half-quadratic (HQ) theory, maximizing the aforementioned correlation entropy is equivalent to minimizing the corresponding weighted quadratic loss function, thus obtaining the denoising loss function based on maximizing the correlation entropy, which is defined as:

[0061] ;

[0062] in, This is a noise reduction loss function based on maximizing the relevant entropy; For the first The weights corresponding to each training sample.

[0063] This can be considered as the reconstruction error for each sample. Weights were applied For samples with large reconstruction errors, their corresponding weights... The weights are relatively small, thus reducing the impact of these samples on the training process; for samples with smaller reconstruction errors, their weights are... The value is close to 1. According to the internal priority principle, the model tends to learn patterns from clean data first, meaning that samples with larger reconstruction errors usually contain more noise. Therefore, the loss function based on relevance entropy is equivalent to assigning smaller weights to noisier samples during the learning process. This adaptive weighting method can be seen as an automatic "soft" selection of samples, weakening the impact of noisier samples on the training process, enabling the model to resist noise interference and focus on extracting the main features of the data during learning.

[0064] Based on the above analysis, the information-theoretic learning-based denoising loss function used for wind turbine data denoising can guide the model to restore noisy data to cleaner data without using clean wind turbine data during the training process.

[0065] To address the characteristics of wind turbine data during training, which may contain Gaussian and non-Gaussian noise and are typically complex, this invention introduces a feature regularization strategy in the latent feature space. This regularization ensures a more uniform distribution of the latent representations of the samples within the feature space. This feature uniformity constraint enhances the expressive power of the data features, enabling the model to not only suppress noise but also preserve the details of different data, thus restoring the true signal as closely as possible. Furthermore, it improves the model's generalization ability, allowing the decoder to learn the mapping from features to clean output over a wider area, ensuring stable noise reduction performance under varying data and noise conditions.

[0066] To achieve the goal of uniform latent feature distribution, a feature regularization strategy is designed from the perspective of information theory learning. Within the framework of information theory learning, when feature values ​​are limited to a finite range, a more uniform distribution results in a higher entropy value. Based on this, feature regularization is achieved by maximizing the entropy of latent features. This embodiment uses Renyi entropy, which guides the distribution of features in the latent space to become more uniform by maximizing the Renyi entropy of latent features. To mitigate the impact of high-noise samples on the training process, the weights of significantly noisy samples are reduced in the regularization strategy to weaken their influence on the model learning process; while cleaner samples are given relatively higher weights to make their distribution more uniform. Simultaneously, the weights in the noise reduction loss function are adjusted... This is incorporated into the feature regularization strategy, thereby constructing a feature regularization loss function based on maximizing Renyi entropy.

[0067] Let the latent features between the encoder and decoder be... (The latent features output by the encoder), whose second-order Renyi entropy is defined as:

[0068]

[0069] in, as a latent feature The second-order Renyi entropy; as a latent feature The approximate probability density function; Approximate probability density function Expected value; It is a logarithmic function.

[0070] The true probability density function Unable to obtain, this embodiment utilizes discrete features. To estimate the distribution of features, where For the first The latent features corresponding to each training sample. This method uses kernel density estimation to approximate the distribution. In modeling feature distribution, to enhance the impact of relatively clean data on the learning process and mitigate the interference of noisy data, a weighted processing of samples is introduced. For latent features... Its approximate probability density estimate is:

[0071]

[0072] in, as a latent feature Approximate probability density estimate; and For indexing, For the first The weight of each sample is assigned; the cleaner the sample, the larger the weight. The Renyi entropy is calculated using the estimated probability density, resulting in the sample-weighted Renyi entropy:

[0073]

[0074] ;

[0075] in, as a latent feature The Renyi entropy based on sample weighting.

[0076] To achieve uniform feature distribution, it is necessary to maximize the Renyi entropy. This embodiment transforms the entropy maximization problem into a minimization problem, thereby defining the feature regularization loss function based on maximizing the Renyi entropy:

[0077]

[0078] The denoising loss function based on maximizing relevance entropy and the feature regularization loss based on maximizing Renyi entropy are weighted and combined to obtain the total loss function used for training, such as... Figure 2 As shown, that is:

[0079]

[0080] in, This is the total loss function for the autoencoder model; Here, represents the weighting coefficients, which are hyperparameters that control the proportion of the two losses during training. By minimizing the total loss function, the parameters of the autoencoder model are iteratively updated. and The training continues until the preset number of iterations is reached or the model convergence condition is met, at which point the trained autoencoder model is obtained. Specifically, the training process does not require the use of clean, noise-free fan data.

[0081] S4. Using historical wind turbine data as training samples, the total loss function of the autoencoder model is constructed by weighted combination of the noise reduction loss function and the feature regularization loss function. The autoencoder model is then iteratively trained using the training samples themselves as the reconstruction target.

[0082] S5. Input the wind turbine data to be denoised into the trained autoencoder model for processing to obtain the denoised wind turbine data.

[0083] The noisy fan data is distorted due to factors such as electromagnetic interference, environmental interference, and pulse interference. The noise-reduced fan data is used for subsequent analysis tasks such as condition monitoring, fault diagnosis, and operation optimization.

[0084] Example 2:

[0085] To further verify the effectiveness of the method of the present invention in noise reduction of real wind turbine data, this embodiment is further verified based on real operating data recorded by the wind turbine monitoring and data acquisition system (SCADA):

[0086] The data originates from an onshore wind farm in Portugal. Relevant data has been published in the literature Christian Gück, Cyriana MA Roelofs, Stefan Faulstich, CARE to Compare: A Real-World Benchmark Dataset for Early Fault Detection in Wind Turbine Data, Data, 2024, 9, 138. The wind turbine data includes multiple monitoring parameters reflecting the turbine's operating status, including wind speed, wind direction, ambient temperature, turbine operating temperature, and power. A continuous time series data segment representing the normal operating status of the wind turbine is selected; this data is used by default as the raw, noise-free data (e.g., ...). Figure 4 The original data shown in the figure is used as a reference for subsequent evaluation of the noise reduction effect.

[0087] Gaussian noise and impulse noise are superimposed on the original noise-free data to simulate the random noise and abnormal disturbances that wind turbine data may encounter during actual data acquisition, thus constructing noisy data samples. Specifically, the noise superposition process includes two types: Gaussian noise and impulse noise. Gaussian noise is superimposed by adding zero-mean random noise to each feature dimension of the original data; impulse noise is used to simulate non-Gaussian noise interference that may occur during wind turbine operation. The superposition method involves randomly selecting a portion of the data with a certain probability and applying a disturbance with an amplitude significantly larger than the normal fluctuation range to these data.

[0088] In this embodiment, the disturbance amplitude is set to 20 times the standard deviation of the original data. By doing so, while maintaining the overall structural characteristics of the original wind turbine operating data, mixed Gaussian and non-Gaussian noise is introduced to construct noisy data that more closely resembles the real wind turbine data acquisition environment (e.g., ...). Figure 4 (as shown in the noisy data), and the noisy data is used as the input data for the model.

[0089] Noisy data is randomly divided into training and testing data. The training data is used to train the autoencoder model, and the testing data is used to verify the model's denoising performance. Only the noisy training data is used during model training; the corresponding noise-free data is not used. After model training is complete, the noisy testing data is input into the trained autoencoder model to obtain the corresponding model output. The model output corresponding to the test sample is compared with its corresponding original noise-free data. The mean squared error (MSE) is used as the evaluation index for denoising performance, measuring the degree of difference between the model output and the original noise-free data. The smaller the value of this index, the better the denoising effect of the model. In terms of model configuration, a cascaded encoder and decoder structure is used to construct the autoencoder model for wind turbine data denoising. In this embodiment, the encoder uses three fully connected layers, and the decoder uses three fully connected layers. Both the encoder and decoder use ReLU as the non-linear activation function.

[0090] To further verify the beneficial effects of the method of the present invention, this embodiment compares the experimental results with those of replacing only the total loss function with the loss function of mean squared error. The comparison results are shown in Table 1 below:

[0091] Table 1:

[0092]

[0093] The total loss function employs mean squared error (MSE) to calculate the MSE based on both the output and input data when calculating the total loss. In summary, the method of this invention demonstrates significant advantages in wind turbine data denoising under complex noise environments. By introducing a denoising loss function based on maximizing correlation entropy, the method maintains good denoising performance even when the data contains non-Gaussian noise, exhibiting stronger noise robustness compared to traditional methods using MSE as the loss function. Furthermore, the method of this invention does not require clean, noise-free data during model training; it can complete model training using only noisy data, overcoming the limitation of difficulty in obtaining clean data in actual wind turbine operation scenarios, thereby improving the applicability and practical value of the method in real engineering environments.

[0094] The noise reduction effect of traditional methods (using the total loss function of mean squared error) is similar to that of the original data residuals. Figure 4 As shown. The denoising effect of the method of this invention (using a denoising loss function + feature regularization loss function) and the residual of the original data are as follows. Figure 4 As shown.

[0095] The above embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims.

Claims

1. A robust noise reduction method for wind turbine data based on information theory learning, characterized in that, Includes the following steps: S1. Obtain wind turbine data under historical wind turbine operating conditions; S2. Construct an autoencoder model that includes a cascaded encoder and decoder; S3. Construct a noise reduction loss function based on maximizing the relevance entropy and a feature regularization loss function based on maximizing the Renyi entropy, respectively. The feature regularization loss function based on maximizing Renyi entropy is set according to the following formula: in, The feature regularization loss function; It is a logarithmic function; For indexing; For the first The weights corresponding to each training sample; and The first The and the first The latent features output by the encoder after processing each training sample; The square of the Euclidean distance between the vectors; For indexing; For the first The weights corresponding to each training sample; The number of training samples; and The first Each training sample and its corresponding model output; This refers to the preset kernel function bandwidth parameters; It is an exponential function; S4. Directly use the historical wind turbine data obtained as training data, and construct the total loss function of the autoencoder model by weighted combination of the noise reduction loss function and the feature regularization loss function. Use the training data itself as the reconstruction target to iteratively train the autoencoder model. S5. Input the wind turbine data to be denoised into the trained autoencoder model for processing to obtain the denoised wind turbine data.

2. The robust noise reduction method for wind turbine data based on information theory learning according to claim 1, characterized in that, Step S1 specifically involves: Multiple operating parameters of historical wind turbine operation status are collected. The operating parameters collected at the same time are combined to construct a single sample of wind turbine data. Wind turbine data samples from consecutive time points constitute wind turbine data. The obtained wind turbine data is standardized to obtain standardized wind turbine data, which is used as training data for the model.

3. The robust noise reduction method for wind turbine data based on information theory learning according to claim 2, characterized in that: The operating parameters include wind speed, wind direction, ambient temperature, unit operating temperature, and power.

4. The robust noise reduction method for wind turbine data based on information theory learning according to claim 1, characterized in that: Both the encoder and decoder are composed of multiple fully connected layers connected in sequence. The encoder receives the input fan data, maps it through the multi-layer fully connected layers of the encoder, and outputs latent features. The decoder receives the latent features and outputs noise-reduced fan data.

5. The robust noise reduction method for wind turbine data based on information theory learning according to claim 1, characterized in that: The noise reduction loss function based on maximizing the correlation entropy is set according to the following formula: ； in, Here is the noise reduction loss function; For indexing; The number of training samples; and The first Each training sample and its corresponding model output; For the first The weights corresponding to each training sample; This refers to the preset kernel function bandwidth parameters; The square of the Euclidean distance between the vectors; It is an exponential function.

6. The robust noise reduction method for wind turbine data based on information theory learning according to claim 1, characterized in that: When iteratively training the autoencoder model, only noisy wind turbine data is used as training data, and the training data itself is used as the reconstruction target. There is no need to use noisy wind turbine data or any labeled data.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images