A method for evaluating the health status of large components in the nacelle of a wind turbine under blade icing conditions

Abstract

The present invention provides a method for evaluating the health status of large components in a wind turbine nacelle under a blade icing state. It relates to the technical field of wind turbine health status evaluation, and in particular to the health status evaluation technology for large components in a wind turbine nacelle under a blade icing state. Through feature extraction and time alignment processing of SCADA data, soundprints, and vibration signals of large components in the wind turbine nacelle under cold weather conditions, a multi-dimensional comprehensive feature matrix is ​​constructed; combining meteorological parameters and blade icing dynamic characteristics, a Transformer network is used to realize blade icing state identification, and a comprehensive feature matrix under blade icing state is output; a Transformer-KAN model is innovatively constructed, and the Transformer network is used to extract the health characteristics of large components in the nacelle under a blade icing state, and the health status classification is completed through the KAN network. The present invention integrates multi-source data, uses deep learning and feature extraction technology, accurately evaluates the impact of blade icing on component health, and provides a basis for operation and maintenance decisions.

Description

Technical Field

[0001] The present invention relates to the field of wind power generation technology, and more particularly to a method for evaluating the health of large components in a wind turbine nacelle when blades are icing. Specifically, the method collects data from various sensors and combines it with data processing techniques to comprehensively assess the health of large components in the wind turbine nacelle. This method can promptly predict and diagnose wind turbine failures or performance degradation in the presence of blade icing, providing effective maintenance decision support. Background Art

[0002] With the widespread deployment of wind turbines in cold regions, blade icing in low-temperature weather conditions has become an increasingly prominent problem. Blade icing not only significantly reduces turbine efficiency but can also cause turbine failure. Therefore, accurately assessing the health of wind turbines in the presence of blade icing and taking timely measures to prevent significant losses have become key technical challenges that the wind power industry urgently needs to overcome.

[0003] Currently, most methods for assessing the health status of key components of wind turbines rely on a single data source, or only integrate acoustic and vibration signals, failing to effectively integrate with other data in the SCADA system. For example, vibration sensors can detect vibration changes caused by blade icing, but their sensitivity and accuracy are easily affected by factors such as wind speed and rotation speed in complex meteorological conditions, resulting in deviations in the diagnosis of icing conditions. While data such as wind speed and unit power in the SCADA system can reflect the overall operation of the unit, they cannot directly reflect the icing condition on the blade surface, and the accuracy of their data may also be affected under extreme climatic conditions. Therefore, existing methods fail to effectively combine multiple data sources and are unable to comprehensively and accurately assess the impact of blade icing on the health status of key components of wind turbines.

[0004] Most of the current health status assessment models are based on data analysis under conventional meteorological conditions, and lack in-depth research on the health status of large nacelle components under cold meteorological conditions, especially under blade icing conditions. The impact of blade icing on wind turbine operation is not limited to the blades themselves, but may also have a negative impact on other important components in the nacelle, such as gearboxes and generators. However, most existing studies only focus on the blade icing phenomenon itself, and rarely involve comprehensive health assessments of large nacelle components under icing conditions. Under blade icing conditions, traditional health assessment models have limitations, mainly reflected in the failure to fully integrate multi-feature data and the failure to effectively integrate the complex relationship between multi-source information such as soundprints and vibrations and large nacelle components, resulting in limited accuracy and credibility of the evaluation results. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for evaluating the health status of large components of a wind turbine nacelle when blades are in an icing state, so as to solve the problems raised in the above background technology.

[0006] To address the above technical issues, the present invention proposes the following technical solution: a method for evaluating the health status of large components in a wind turbine nacelle under blade icing conditions. The health status evaluation method of the present invention primarily involves collecting and fusing data from large components in the nacelle under cold weather conditions, automatically identifying the blade icing condition, and establishing a health status evaluation model for blades in this icing condition. The method includes the following steps:

[0007] S1 First, obtain the SCADA (Supervisory Control and Data Acquisition) data, voiceprints, and vibration signals of the large components of the nacelle (main shaft, gearbox, generator, blades, etc.) of the wind turbine under cold weather conditions. Secondly, perform data cleaning and standardization on the SCADA data, including denoising, standardization, missing data filling, and outlier detection to ensure data quality and provide a reliable basis for subsequent analysis. Then, perform denoising, standardization, and feature extraction on the voiceprints and vibration signals, mainly extracting frequency domain features, time domain features, and statistical features to reflect the operating characteristics of the wind turbine under different working conditions and extract key information from the data. Finally, perform feature fusion on the processed SCADA data, voiceprints, and vibration signals, mainly using fusion methods such as weighted averaging and dimensionality reduction mechanisms, and fully combine information from different data sources to generate a comprehensive feature matrix of the same time scale, thereby improving the accuracy and robustness of subsequent model predictions.

[0008] S2 inputs the comprehensive feature matrix obtained by S1, meteorological data (such as temperature, humidity, wind speed, etc.) and blade surface icing characteristics into the pre-trained blade icing identification model to obtain the blade icing identification result. The blade icing identification model is based on the Transformer network, and uses the comprehensive feature matrix obtained in S1, combined with meteorological conditions and blade surface icing characteristics, to effectively distinguish between the blade icing state and the normal working state. Specifically, the model learns the feature distribution under each state based on the comprehensive feature matrix obtained in S1, so as to determine whether there is icing on the blade. The output of the model is the judgment result of the blade icing state and the comprehensive feature matrix under the blade icing state, and provides a basis for subsequent health status assessment.

[0009] S3 inputs the comprehensive feature matrix of the blade icing state obtained in S2 into the Transformer-KAN model for health status assessment. The Transformer-KAN model classifies the health status of large components in the wind turbine nacelle, wherein the health status characteristics of the large components in the nacelle are extracted based on the Transformer network, and the extracted health status characteristics are specifically classified based on the KAN (Kolmogorov-Arnold Networks) model. And it evaluates whether there is a risk of failure of large components in the nacelle due to blade icing. Subsequently, the model assigns a health status score to each large component in the nacelle. The scoring result reflects the health status of the large components in the nacelle under the blade icing state. The scoring levels can be divided into health, mild failure, moderate failure and severe failure. The evaluation results provide data support for subsequent operation and maintenance decisions.

[0010] Furthermore, the SCADA data in S1 mainly includes gearbox oil temperature, gearbox bearing temperature, average wind speed, generator bearing temperature, generator power, speed, etc.; the soundprint and vibration signals include main shaft vibration, gearbox vibration, generator vibration, cabin soundprint, blade soundprint, etc.

[0011] Furthermore, frequency domain features, time domain features, and statistical features are extracted from the voiceprint and vibration signals in S1. Time domain features include mean, root mean square, kurtosis, etc., which are used to reflect the volatility, amplitude, and variation of the signal; frequency domain features include spectrum, main frequency, frequency bandwidth, etc., which can reveal the vibration mode, frequency components, and potential fault frequencies of the device; and statistical features include peak factor, frequency domain entropy, signal energy, etc., which can reveal the stability, complexity, and potential fault modes of the device.

[0012] Furthermore, the features from different signal sources (SCADA data, voiceprints, and vibration signals) in S1 are fused to improve the accuracy and robustness of the model prediction. Common fusion methods include weighted averaging and KPCA (Kernel Principal Component Analysis) dimensionality reduction. Weighted averaging generates a comprehensive feature vector by assigning weights to different signal sources and combining their respective importance. It is suitable for situations where the contribution of signal sources is uneven; while KPCA dimensionality reduction maps data from high-dimensional space to low-dimensional space through nonlinear mapping, extracts the most representative features, reduces redundant information, and improves computational efficiency. These fusion methods can effectively combine the advantages of multi-source data, provide comprehensive and accurate feature representation, and enhance the performance of subsequent analysis and prediction.

[0013] Furthermore, the SCADA data, voiceprint, and vibration signals from the feature fusion in S1 are time-aligned to ensure their consistency in the temporal dimension. This process is achieved by synchronizing the timestamps of different data sources, thereby generating time-aligned multi-source fusion data, providing a unified time base for subsequent deep learning models.

[0014] Furthermore, the icing characteristics of the blade surface in S2 are obtained through the following technical means: the ice layer type is divided into frost ice, transparent ice or mixed ice using infrared thermal imager and visible light camera images through convolutional neural network (CNN); the ice layer thickness is directly measured by the ultrasonic sensor array on the leading edge of the blade, or dynamically estimated by combining environmental parameters and Macklin ice growth model; the infrared reflectivity is extracted by near-infrared spectrometer to obtain the mean and standard deviation of the reflectivity in the 800-1200nm band; the acoustic characteristics are analyzed from the soundprint signal to obtain the energy attenuation rate and spectral entropy in the 2-5kHz frequency band; the surface texture is quantified by the root mean square value of the high-frequency component (5-10kHz) of the vibration signal and the wavelet packet energy entropy.

[0015] Furthermore, when building the blade icing identification model described in S2, the core of the Transformer architecture is the multi-head self-attention mechanism, which can process different parts of the input data in parallel, thereby capturing long-range dependencies and local features in the data. By decomposing the input data into multiple "attention heads," each focusing on a different aspect of the input sequence, the model can more comprehensively understand the complex relationships between the integrated feature data. In this way, the model can efficiently process the integrated feature matrix and extract feature information related to the blade icing state.

[0016] To preserve the positional information in the input sequence, the model introduces positional encoding. Positional encoding adds a fixed vector to each position, enabling the model to perceive positional relationships within the sequence. The introduction of positional encoding ensures that the model preserves temporal order when processing sequential data, which is particularly important for dynamic monitoring of blade icing conditions.

[0017] In addition to the multi-head self-attention mechanism, each layer of the encoder also includes a feed-forward neural network (FFN). The FFN performs the same fully connected operation on the feature vector at each position to further extract high-level feature representations. Through the FFN, the model can perform nonlinear transformations on the output of the multi-head self-attention mechanism, thereby enhancing the model's expressive power.

[0018] During model training, the cross-entropy loss function is used to measure the difference between the predicted results and the true labels. This function effectively evaluates the model's classification performance and optimizes model parameters by minimizing the difference between the predicted probability distribution and the true labels. To efficiently update model parameters, the Adam (Adaptive Moment Estimation) optimizer is used. The Adam optimizer combines the advantages of momentum and adaptive learning rates, enabling rapid convergence and stable training. By adjusting the learning rate and hyperparameters, the model can achieve high accuracy within a limited training time.

[0019] Through this design, the Transformer network can effectively process the comprehensive feature matrix and effectively distinguish between blade icing and normal operating conditions. The model uses a self-attention mechanism to learn the relationships between comprehensive feature data, combines positional encoding to preserve sequence information, and extracts high-level features through a feedforward neural network. This architecture not only improves the model's expressiveness but also enhances its adaptability to complex data.

[0020] Furthermore, in S3, the Transformer network is used to extract health status features from the comprehensive feature matrix. The Transformer encoder captures global dependencies in the data through a multi-head self-attention mechanism and FFN, generating context-sensitive feature representations. The multi-head self-attention mechanism divides the input data into multiple "heads" for parallel computation. The feature representations learned by each head are then concatenated and linearly transformed, thereby capturing features from different subspaces in the multi-source fused data. The FFN further processes these feature representations, increasing the model's nonlinear capabilities through two linear transformations and a nonlinear activation function. Furthermore, layer normalization is used to reduce internal covariate shift and accelerate model convergence, while a dropout layer is used to prevent overfitting. To extract global features and reduce feature dimensionality, the Transformer output feature maps are pooled into fixed-size feature vectors through average pooling. The pooled feature maps are flattened into one-dimensional vectors for input into the subsequent fully connected layers.

[0021] Furthermore, based on the features extracted by the Transformer network in S3, the KAN model is introduced to classify health status. The KAN model uses basic linear transformations and piecewise polynomial basis functions (B-spline) for nonlinear mapping, improving the model's expressiveness and interpretability. The KAN model classifies the extracted health status features into healthy, mildly faulty, moderately faulty, and severely faulty. In the KAN model, input features are first mapped to the output space through a linear transformation, increasing the model's linear expressiveness.

[0022] The input features are mapped to multiple piecewise polynomial spaces through B-spline basis functions. Each piecewise polynomial can capture the nonlinear relationship of the input data in different intervals.

[0023] By combining basic activation functions and piecewise polynomial basis functions, the input features are nonlinearly mapped to generate a richer feature representation. Ultimately, the KAN model outputs a vector, each element of which represents the predicted probability that the sample belongs to the corresponding health state.

[0024] This results in the output vector of the Transformer-KAN model, where each element represents the predicted probability that the sample belongs to the corresponding health state.

[0025] Furthermore, to comprehensively evaluate the performance of the Transformer-KAN model in S3, metrics such as accuracy, precision, recall, F1 score, and confusion matrix were used. These metrics provide a comprehensive assessment of the model's performance in blade ice detection and health status classification. Learning rate scheduling and regularization techniques, combined with optimization using the Simulated Firefly Optimization Algorithm (SFOA), further optimize the model training process, prevent overfitting, and accelerate convergence.

[0026] Furthermore, the health status of each sample is determined based on the predicted probabilities output by the Transformer-KAN model described in S3. By setting thresholds or using classification strategies (such as maximum probability selection), samples are classified into specific health status categories, including healthy, minor fault, moderate fault, and severe fault. This classification result provides a scientific basis for operation and maintenance decisions, ensuring the efficient operation and maintenance of wind turbines.

[0027] The beneficial effects of adopting this patent are: by fusing SCADA data with voiceprints and vibration signals, the impact of blade icing on large components in the wind turbine nacelle is fully considered, thereby improving the accuracy and reliability of health status assessment. The introduction of the Transformer network and KAN model can effectively process complex multi-source data and improve the expressiveness and interpretability of the model. Time alignment and feature extraction techniques are used to ensure data consistency and model robustness. By optimizing algorithms and performance evaluation indicators, the training efficiency and classification performance of the model are further improved. It provides a scientific basis for operation and maintenance decisions of wind turbines in cold weather conditions, reduces the risk of failures caused by blade icing, and ensures the safe and efficient operation of the unit.

[0028] The above description is only an overview of the technical solution of this application. To better understand the technical means of this application, which can be implemented in accordance with the contents of the description, and to make the above and other purposes, features and advantages of this application more obvious and easy to understand, the following specific embodiments of this application are specifically described. It should be understood that the above general description and the detailed description below are merely exemplary and do not limit the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] To more clearly illustrate the technical solutions implemented in the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. Those skilled in the art can derive other drawings based on these drawings without inventive effort.

[0030] Figure 1 This is a business execution flow chart of the method for evaluating the health status of large components in a wind turbine nacelle under blade icing conditions according to the present invention;

[0031] Figure 2 This is a flowchart of the business execution of data preprocessing and feature fusion described in the present invention;

[0032] Figure 3 This is a business execution flow chart for the health status evaluation of large components of a wind turbine based on the Transformer-KAN model according to the present invention; DETAILED DESCRIPTION

[0033] The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0034] The health status evaluation method of large components in the wind turbine nacelle under blade icing condition, such as Figure 1 As shown in the figure, the health status evaluation method for large components of wind turbine nacelles under blade icing conditions accurately identifies blade icing conditions and evaluates the health status of large components of wind turbine nacelles under blade icing conditions through data collection, preprocessing, fusion and analysis, combined with a deep learning model, to ensure the safe and efficient operation of the unit under cold weather conditions. The specific implementation plan is as follows:

[0035] S1 Figure 2As shown, first, the SCADA data, voiceprints and vibration signals of the major components of the wind turbine nacelle (main shaft, gearbox, generator, blades, etc.) under cold weather conditions are obtained. Secondly, the SCADA data is cleaned and standardized, specifically including denoising, standardization, missing data filling and outlier detection, to ensure data quality and provide a reliable basis for subsequent analysis. Then, the voiceprints and vibration signals are denoised, standardized and feature extracted, mainly extracting frequency domain features, time domain features and statistical features to reflect the operating characteristics of the wind turbine under different working conditions, and extracting key information from the data. Finally, the processed SCADA data, voiceprints and vibration signals are feature fused, mainly using fusion methods such as weighted averaging and dimensionality reduction mechanisms, and fully combining information from different data sources to generate a comprehensive feature matrix of the same time scale, thereby improving the accuracy and robustness of subsequent model predictions.

[0036] S2 inputs the comprehensive feature matrix obtained by S1, meteorological data (such as temperature, humidity, wind speed, etc.) and blade surface icing characteristics into the pre-trained blade icing identification model to obtain the blade icing identification result. The blade icing identification model is based on the Transformer network, and uses the comprehensive feature matrix obtained in S1, combined with meteorological conditions and blade surface icing characteristics, to effectively distinguish between the blade icing state and the normal working state. Specifically, the model learns the feature distribution under each state based on the comprehensive feature matrix obtained in S1, so as to determine whether there is icing on the blade. The output of the model is the judgment result of the blade icing state and the comprehensive feature matrix under the blade icing state, and provides a basis for subsequent health status assessment.

[0037] S3 Figure 3 As shown, the comprehensive feature matrix under the blade icing state obtained in S2 is input into the Transformer-KAN model for health status assessment. The Transformer-KAN model classifies the health status of large components in the wind turbine nacelle, wherein the health status features of the large components in the nacelle are extracted based on the Transformer network, and the extracted health status features are specifically classified based on the KAN model. And it is evaluated whether there is a risk of failure of large components in the nacelle due to blade icing. Subsequently, the model assigns a health status score to each large component in the nacelle. The scoring result reflects the health status of the large components in the nacelle under the blade icing state. The scoring level can be divided into health, mild fault, moderate fault and severe fault, etc. The evaluation result provides data support for subsequent operation and maintenance decisions.

[0038] Furthermore, the sampling interval of the SCADA data in S1 is 1 minute, which mainly includes gearbox oil temperature, gearbox drive end bearing temperature, gearbox non-drive end bearing temperature, average wind speed, generator front bearing temperature, generator rear bearing temperature, generator stator winding temperature, average generator power, etc.; the sampling frequency of sound, vibration and speed signals is 12.8kHz, which mainly includes: horizontal vibration of the main shaft front bearing, horizontal vibration of the main shaft rear bearing, axial vibration of the gearbox low-speed shaft, axial vibration of the gearbox high-speed shaft, vertical vibration of the generator drive end, vertical vibration of the generator non-drive end, cabin sound pattern, blade sound pattern, etc.

[0039] Furthermore, frequency domain features, time domain features, and statistical features are extracted from the voiceprint and vibration signals in S1. Time domain features include mean, root mean square, kurtosis, etc., which are used to reflect the volatility, amplitude, and variation of the signal; frequency domain features include spectrum, main frequency, frequency bandwidth, etc., which can reveal the vibration mode, frequency components, and potential fault frequencies of the device; and statistical features include peak factor, frequency domain entropy, signal energy, etc., which can reveal the stability, complexity, and potential fault modes of the device.

[0040] Furthermore, the features from different signal sources (SCADA data, voiceprints, and vibration signals) in S1 are fused to improve the accuracy and robustness of the model prediction. Common fusion methods include weighted averaging and KPCA dimensionality reduction. Weighted averaging generates a comprehensive feature vector by assigning weights to different signal sources and combining their respective importance. It is suitable for situations where the contribution of signal sources is uneven. KPCA dimensionality reduction maps data from high-dimensional space to low-dimensional space through nonlinear mapping, extracts the most representative features, reduces redundant information, and improves computational efficiency. These fusion methods can effectively combine the advantages of multi-source data, provide comprehensive and accurate feature representation, and enhance the performance of subsequent analysis and prediction.

[0041] Furthermore, the SCADA data, voiceprint, and vibration signals from the feature fusion in S1 are time-aligned to ensure their consistency in the temporal dimension. This process is achieved by synchronizing the timestamps of different data sources, thereby generating time-aligned multi-source fusion data, providing a unified time base for subsequent deep learning models.

[0042] Furthermore, the main characteristics of blade surface icing in S2 are: uneven distribution of ice on the leading edge and upper surface of the blade. Morphologically, it is classified by CNN as frost ice (loose and porous), transparent ice (dense and smooth), or mixed ice (multi-phase composite). Its formation is synergistically influenced by temperature (-15°C to 0°C), humidity (>85% RH), and wind speed (>4m / s). The ice layer thickness is measured by ultrasonic wave and is 0.5-10mm, or the model estimates the gradient to be 0.1-2mm / h, resulting in a localized blade mass increase of 5%-30%. The dynamic growth of the ice layer causes abnormal fluctuations in infrared reflectivity (decreases of 20%-40%), with the mean value dropping to 40%-60% (normal 70%-85%) and the standard deviation increasing to 8%-12%. This is accompanied by a shift in the acoustic signal's spectral characteristics (high-frequency energy attenuation >15dB), with energy attenuation in the 2-5kHz band exceeding 15% and a spectral entropy of 1.2-1.8 (normal 0.6-1.0). Surface roughness increases significantly (Ra >50μm), with the root mean square vibration value in the 5-10kHz band reaching 0.8-1.5g (normal 0.3-0.6g), and wavelet packet energy entropy ranging from 4.5-6.0 (normal 2.5-3.5). These features are then time-aligned and fed into a Transformer network to support the construction of a blade ice identification model.

[0043] Furthermore, when building the blade icing recognition model described in S2, the core of the Transformer architecture is the multi-head self-attention mechanism, which can process different parts of the input data in parallel, thereby capturing long-range dependencies and local features in the data. By decomposing the input data into multiple "attention heads," each focusing on a different aspect of the input sequence, the model can more comprehensively understand the complex relationships between multi-source data. Specifically, the calculation formula of the multi-head self-attention mechanism is:

[0044] MultiHead(Q,K,V)=Concat(head1,head2,…,head h )W O

[0045] in,

[0046] head i =Attention(QW i Q ,KW i K ,VW i V )

[0047]

[0048] Here, Q, K, and V represent the query, key, and value matrices respectively. k is the dimension of the key, h is the number of attention heads, and W O is the output projection matrix. In this way, the model can efficiently process the comprehensive feature data and extract feature information related to the blade icing state.

[0049] In order to preserve the position information in the input sequence, the model introduces position encoding. Position encoding adds a fixed vector to each position, allowing the model to perceive the position relationship in the sequence. The calculation formula of position encoding is:

[0050]

[0051] Among them, pos represents position, i represents dimension, d model is the dimension of the model. The introduction of position encoding ensures that the model can retain the temporal order information when processing sequence data, which is particularly important for the dynamic monitoring of blade icing status.

[0052] In each layer of the encoder, in addition to the multi-head self-attention mechanism, a feed-forward neural network is also included. FFN performs the same fully connected operation on the feature vector at each position to further extract high-level feature representations. Its calculation formula is:

[0053] FFN(x)=max(0,xW1+b1)W2+b2

[0054] Among them, W1, W2, b1, and b2 are network parameters. Through FFN, the model can perform nonlinear transformations on the output of the multi-head self-attention mechanism, thereby enhancing the model's expressive power.

[0055] During model training, the cross-entropy loss function is used to measure the difference between the predicted results and the true labels. The cross-entropy loss function can effectively evaluate the classification performance of the model and optimize the model parameters by minimizing the difference between the predicted probability distribution and the true label. The specific formula is:

[0056]

[0057] Among them, y is the true label, is the model's predicted output. To efficiently update model parameters, the Adam optimizer is used. The Adam optimizer combines the advantages of momentum and adaptive learning rates, enabling rapid convergence and stabilizing the training process. By adjusting the learning rate and hyperparameters, the model can achieve high accuracy within a limited training time.

[0058] Through this design, the Transformer network can effectively process the comprehensive feature matrix and effectively distinguish between blade icing and normal operating conditions. The model uses a self-attention mechanism to learn the relationships between comprehensive feature data, combines positional encoding to preserve sequence information, and extracts high-level features through a feedforward neural network. This architecture not only improves the model's expressiveness but also enhances its adaptability to complex data.

[0059] Furthermore, the time-aligned multi-source fusion data in S3 is subjected to feature extraction through a Transformer network. The Transformer encoder captures global dependencies in the data through a multi-head self-attention mechanism and FFN, generating context-sensitive feature representations. The multi-head self-attention mechanism divides the input data into multiple "heads" for parallel computation. The feature representations learned by each head are then concatenated and linearly transformed, thereby capturing features from different subspaces within the multi-source fusion data. The FFN further processes these feature representations, increasing the model's nonlinear capabilities through two linear transformations and a nonlinear activation function. Furthermore, layer normalization is used to reduce internal covariate shift and accelerate model convergence, while a Dropout layer is used to prevent overfitting. To extract global features and reduce feature dimensionality, the Transformer output feature maps are pooled into fixed-size feature vectors through average pooling. The pooled feature maps are flattened into one-dimensional vectors for input into the subsequent fully connected layers.

[0060] Furthermore, based on the features extracted by the Transformer network in S3, the KAN model is introduced to perform health status classification. The KAN model uses basic linear transformations and B-spline basis functions for nonlinear mapping, improving the model's expressiveness and interpretability. The KAN model classifies the extracted health status features into healthy, mild fault, moderate fault, and severe fault.

[0061] In the KAN model, the input features are first mapped to the output space through linear transformation to increase the linear expression ability of the model. The calculation formula of linear transformation is:

[0062] BaseLinear=Wx+b

[0063] Among them, x is the input feature, W is the basic weight matrix, and b is the bias vector.

[0064] The input features are mapped to multiple piecewise polynomial spaces through B-spline basis functions. Each piecewise polynomial can capture the nonlinear relationship of the input data in different intervals. The calculation formula of B-spline basis function is:

[0065]

[0066] Among them, x is the input feature, c i are the coefficients of the piecewise polynomial, B i (x) is the i-th B-spline basis function.

[0067] By combining the basic activation function and the piecewise polynomial basis function, the input features are nonlinearly mapped to generate a richer feature representation. Ultimately, the KAN model outputs a vector, each element of which represents the predicted probability of the sample belonging to the corresponding health state. The basic activation function can use a nonlinear function such as SiLU, and the calculation formula is:

[0068] Nonlinear(x)=BaseLinear(Activation(x))+B-spline(x)

[0069] Among them, Activation(x) is a basic activation function, such as SiLU function.

[0070] This results in the output vector of the Transformer-KAN model, where each element represents the predicted probability that the sample belongs to the corresponding health state.

[0071] Furthermore, to comprehensively evaluate the performance of the Transformer-KAN model in S3, metrics such as accuracy, precision, recall, F1 score, and confusion matrix were used. These metrics provide a comprehensive assessment of the model's performance in blade ice detection and health status classification. Learning rate scheduling and regularization techniques, combined with the Firefly Optimization algorithm, further optimize the model training process, prevent overfitting, and accelerate convergence.

[0072] Furthermore, the health status of each sample is determined based on the predicted probabilities output by the Transformer-KAN model described in S3. By setting thresholds or using classification strategies (such as maximum probability selection), samples are classified into specific health status categories, including healthy, minor fault, moderate fault, and severe fault. This classification result provides a scientific basis for operation and maintenance decisions, ensuring the efficient operation and maintenance of wind turbines.

[0073] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some or all of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.

Claims

1. A method for evaluating the health status of large components in a wind turbine nacelle under blade icing conditions. Its features are as follows: including: S1: Acquire SCADA data, soundprints, and vibration signals of large components of the wind turbine nacelle under cold weather conditions; perform data cleaning and standardization on the SCADA data, and perform denoising, standardization, and feature extraction on the soundprints and vibration signals; perform feature fusion on the processed SCADA data, soundprints, and vibration signals to generate a comprehensive feature matrix with the same time scale; S2: Inputting the comprehensive feature matrix obtained in S1, the pre-acquired meteorological data, and the blade surface icing features into a pre-trained blade icing identification model to obtain a comprehensive feature matrix under the blade icing state; the blade icing identification model is constructed based on a Transformer network; S3: The comprehensive feature matrix of the blade icing state obtained in S2 is input into the Transformer-KAN model for health status assessment; the Transformer-KAN model classifies the health status of large components in the wind turbine nacelle, wherein the health status features of the large components in the nacelle are extracted based on the Transformer network, and the extracted health status features are specifically classified based on the KAN model, and an assessment is made as to whether there is a risk of failure of the large components in the nacelle due to blade icing.

2. The method for evaluating the health status of large components in a wind turbine nacelle under blade icing conditions according to claim 1 is characterized as follows: the SCADA data in step S1 includes gearbox oil temperature, gearbox bearing temperature, average wind speed, generator bearing temperature, generator power, and speed; and the soundprint and vibration signals include main shaft vibration, gearbox vibration, generator vibration, nacelle soundprint, and blade soundprint.

3. The method for evaluating the health status of large components in a wind turbine nacelle under blade icing conditions according to claim 1 is characterized in that: in step S1, the SCADA data is cleaned and standardized, including denoising, standardization, missing data filling, and outlier detection; and the features extracted from the voiceprint and vibration signals include time domain features, frequency domain features, and statistical features.

4. The method for evaluating the health status of large components in a wind turbine nacelle under blade icing conditions according to claim 1, characterized in that: in S1, the feature fusion method includes weighted averaging and KPCA (Kernel Principal Component Analysis) dimensionality reduction; and time alignment processing is performed on the SCADA data after feature fusion, as well as the voiceprint and vibration signals, to provide a unified time reference.

5. The method for evaluating the health status of large components of a wind turbine nacelle under blade icing conditions according to claim 1 is characterized by: The meteorological data in step S2 include temperature, humidity, wind speed, air pressure and dew point temperature, which are obtained from the SCADA system or an external meteorological station; the icing characteristics of the blade surface include ice type, thickness, infrared reflectivity, acoustic properties and surface texture, which are obtained through multimodal sensors or physical models.

6. The method for evaluating the health status of large components in a wind turbine nacelle under blade icing conditions according to claim 1, wherein: in S2, during the training of the blade icing identification model based on the Transformer network, a cross-entropy loss function is used to measure the difference between the predicted result and the true label, and the model parameters are updated using the Adam (Adaptive Moment Estimation) optimizer to improve the classification performance of the model.

7. The method for evaluating the health status of large components of a wind turbine nacelle under blade icing conditions according to claim 1, characterized in that: In S3, a Transformer network is introduced to extract the health status features of large cabin components; the Transformer encoder captures global dependencies through a multi-head self-attention mechanism and a feed-forward neural network (FFN), and optimizes feature representation through layer normalization, dropout layers, and average pooling.

8. The method for evaluating the health status of large components of a wind turbine nacelle under blade icing conditions according to claim 1, characterized in that: In S3, the KAN model is introduced to perform health status classification, and nonlinear mapping is performed through linear transformation and piecewise polynomial basis function (B-spline) to classify samples into healthy, mild fault, moderate fault and severe fault categories.

9. The method for evaluating the health status of large components of a wind turbine nacelle under blade icing conditions according to claim 1, characterized in that: In S3, accuracy, precision, recall, F1 score, and confusion matrix are used to evaluate model performance, and the Simulated Firefly Optimization Algorithm (SFOA) is used for optimization.

Images