A wind turbine blade health state evaluation method based on LNN and improved LightGBM

Abstract

The application discloses a wind turbine blade health state evaluation method based on LNN and an improved LightGBM, relates to the technical field of wind turbine health state evaluation, and in particular to the health state evaluation technology of wind turbine blades. The soundprint signal, vibration signal and wind turbine SCADA data of the blade are acquired, and the data is preprocessed; the fusion of the soundprint features and vibration features is realized by adopting an attention mechanism, and a multidimensional feature matrix is constructed in combination with the SCADA data; dynamic feature deep extraction is carried out through a liquid neural network (LNN), so that a high-dimensional feature vector capable of representing the health state of the blade is obtained; and the improved LightGBM model is inputted, so that the health state grade evaluation of the blade is realized. The soundprint, vibration signal and SCADA data are associated and reconstructed, the advantages of the LNN network in capturing dynamic features and the efficient classification ability of the improved LightGBM model are combined, the accurate evaluation of the health state of the blade is realized, and reliable decision-making basis is provided for the maintenance of the wind turbine blade.

Description

Technical Field

[0001] This invention relates to the field of wind turbine health status assessment, and more particularly to a method for assessing the health status of wind turbine blades based on LNN and an improved LightGBM. By fusing acoustic and vibration signal features, combining the advantages of LNN networks in capturing dynamic features with the efficient classification capabilities of LightGBM, this invention can effectively and comprehensively assess the health status of wind turbine blades, providing intelligent auxiliary judgment criteria for wind power operation and maintenance. Background Technology

[0002] As the core component for capturing wind energy, the health of wind turbine blades directly determines the power generation efficiency and operational safety of the unit. Exposed to a complex natural environment for extended periods, blades must withstand multiple factors, including aerodynamic loads, turbulent impacts, sudden temperature changes, dust erosion, and lightning strikes, making them susceptible to crack propagation, leading-edge corrosion, icing, and structural fatigue. Statistics show that blade failures account for more than a quarter of total wind turbine downtime, and repair costs can reach 15% to 20% of the total unit cost. Therefore, accurate monitoring and assessment of blade health is a key requirement for intelligent operation and maintenance of wind farms.

[0003] Current wind turbine blade health status assessment technologies are mainly divided into two categories: traditional assessment methods and data-driven methods. Among traditional assessment methods, visual inspection relies on operator experience, which suffers from high subjectivity, low efficiency, difficulty in detecting internal defects, and inability to operate in adverse weather conditions. Ultrasonic and X-ray inspections require offline operation with the turbine shut down, impacting power generation revenue, and demanding high levels of equipment and personnel skill. Single-sensor monitoring methods are commonly used in data-driven fields. Vibration monitoring involves installing accelerometers to collect blade vibration signals and extracting fault features based on spectral analysis and wavelet transform. However, blade vibration signals are easily interfered with by turbine drivetrain vibration and turbulent disturbances, resulting in a low signal-to-noise ratio and difficulty in distinguishing between normal operating fluctuations and early minor faults. Acoustic monitoring uses microphones to collect acoustic radiation signals during blade operation and uses spectrogram analysis to identify abnormal noise; however, strong environmental noise interference leads to poor assessment stability in complex wind farm environments.

[0004] In recent years, deep learning technology has been initially applied in blade health status assessment, such as CNN models based on vibration signals and LSTM models based on acoustic signals. However, existing deep learning methods still have significant limitations: on the one hand, most models rely only on single vibration or acoustic data, failing to fully utilize the complementary information of the two types of signals. Vibration signals are sensitive to blade structural damage, while acoustic signals can reflect aerodynamic anomalies; a single data source is insufficient to comprehensively characterize the complex failure modes of blades. On the other hand, dynamic changes in operating conditions during blade operation lead to strong non-stationarity and feature drift in acoustic and vibration signals. Existing models are mostly trained based on fixed operating condition data, resulting in insufficient adaptability to changes in operating conditions and poor generalization performance. Furthermore, traditional data-driven methods lack interpretable analysis of fault location and severity, only enabling the classification and judgment of health and faults, which cannot meet the needs of wind farms for precise operation and maintenance at the blade component level.

[0005] The occurrence and evolution of blade failures are the result of the coupled effects of multiple factors, including structural damage and aerodynamic disturbances, and their acoustic and vibration signals contain rich complementary information. Vibration signals can reflect changes in blade structural stiffness and mass distribution, while acoustic signals can reflect aerodynamic anomalies under fluid-structure interaction. Fusion of these two methods can achieve multidimensional complementarity and redundancy suppression of fault characteristics. However, existing fusion methods are mostly simple feature-level splicing, failing to deeply explore the dynamic correlation and spatiotemporal coupling characteristics between acoustic and vibration signals, resulting in poor fusion effects. Therefore, there is an urgent need for a health status assessment method that can deeply fuse acoustic and vibration data and adapt to dynamic changes in operating conditions. Summary of the Invention

[0006] The purpose of this invention is to provide a method for assessing the health status of wind turbine blades based on LNN and an improved LightGBM. By constructing an efficient acoustic and vibration feature fusion model, the method improves the accuracy and stability of early blade health status assessment, providing technical support for predictive maintenance of wind turbine blades.

[0007] To address the aforementioned technical problems, this invention proposes the following technical solution: a method for assessing the health status of wind turbine blades based on LNN and an improved LightGBM model. The health status assessment method of this invention includes the following steps: obtaining an acoustic-vibration fusion feature vector by fusing the acoustic signature and vibration features of the blades; constructing a multi-dimensional feature matrix using SCADA data from the wind turbine; performing deep learning classification of the feature matrix using an LNN network combined with an improved LightGBM model; and outputting the wind turbine blade health status level assessment result. The specific steps are as follows: S1: Acoustic and vibration signals of wind turbine blades are collected using acoustic and vibration sensors. SCADA data of the wind turbine's operating status is obtained from the wind turbine monitoring system. Preprocessing of the acoustic, vibration, and SCADA data is performed, specifically including wavelet threshold denoising and Fourier transform for the blade acoustic signals; low-pass filtering, linear interpolation, and detrending processing for the vibration signals; and trend separation, moving average filtering, outlier detection and correction, interpolation compensation, and normalization processing for the SCADA data.

[0008] S2: Extract features from the preprocessed voiceprint signal and vibration signal to obtain voiceprint features and vibration features respectively. Input the voiceprint features and vibration features into the attention mechanism module to obtain the sound and vibration fusion feature vector. Associate the sound and vibration fusion feature vector with the preprocessed SCADA data to reconstruct a multi-dimensional feature matrix and perform normalization processing. S3: Input the normalized multidimensional feature matrix into the LNN network, and obtain a high-dimensional feature vector that can characterize the health status of the leaf by dynamically extracting features from the multidimensional feature matrix through the LNN network. S4: The improved LightGBM model, through particle swarm optimization and cross-validation, performs deep classification learning on high-dimensional feature vectors to determine the health status level of wind turbine blades, providing a more reliable basis for turbine maintenance.

[0009] Furthermore, the acoustic sensor described in S1 is deployed at the root of the three blades of the wind turbine generator. This location can effectively collect the acoustic signals generated by the blade vibration while reducing the interference of environmental noise. The vibration sensor is deployed at the junction of the hub and the three blades and extends into the blade. It can directly acquire the original signal of the blade structure vibration and improve the signal-to-noise ratio.

[0010] Furthermore, the SCADA data mentioned in S1 includes turbulence intensity, wind speed, blade speed, hub vibration value, generator power, generator speed, pitch shaft rate, and pitch shaft position deviation.

[0011] Furthermore, the voiceprint features and vibration features mentioned in S2 both include time-domain features and frequency-domain features: the time-domain features include root mean square, variance, standard deviation, skewness, and kurtosis; the frequency-domain features include dominant frequency amplitude, bandwidth, spectral entropy, spectral flatness, and spectral roll-off point.

[0012] Furthermore, in the dynamic feature fusion process of acoustic and vibration signals using the attention mechanism described in S2, the acoustic signature features and vibration features are input into the attention mechanism module. By calculating the correlation score between the features and normalizing the correlation score using the Softmax function, different weights can be automatically assigned according to the differences in the contribution of acoustic and vibration signals to fault diagnosis or condition monitoring tasks under different operating conditions. The acoustic signature features and vibration features are multiplied element-wise with their corresponding dynamic weights and then superimposed to obtain a complementary acoustic and vibration fusion feature vector.

[0013] Furthermore, the multidimensional feature matrix described in S2 uses the sampling time point of the SCADA data as a reference to perform spatiotemporal alignment processing on the acoustic-vibration fusion feature vector to ensure that the acoustic-vibration fusion feature vector is consistent with the SCADA data in terms of time scale. The spatiotemporally aligned acoustic-vibration fusion feature vector is then associated and reconstructed with the preprocessed SCADA data to obtain the multidimensional feature matrix.

[0014] Furthermore, the normalization process described in S2 is performed using the Z-score normalization method. Z-score normalization can convert different features into a standard distribution with zero mean and unit variance, effectively eliminating the influence of dimensions between features.

[0015] Furthermore, the LNN network described in S3 uses the local receptive field as the core feature extraction mechanism. For the multi-dimensional feature matrix constructed from the blade acoustic-vibration fusion data, it sets a reasonable receptive field size and step size to accurately capture the dynamic correlation between adjacent time steps. The changes in adjacent moments of the acoustic-vibration fusion data exhibit strong correlations. The local receptive field can adaptively extract local temporal dependencies at different time scales using a sliding window approach, avoiding redundant computations and irrelevant information interference caused by global connections, thus laying the foundation for subsequent feature integration.

[0016] After local feature extraction, the LNN network systematically integrates multi-dimensional local features through fully connected layers. Relying on learnable weight matrices, the fully connected layers map local features scattered across different receptive fields to a unified feature space, achieving cross-temporal information interaction and complementary fusion through matrix operations. This integration mechanism not only fully preserves the temporal specificity of each local feature but also effectively eliminates the information limitations of a single local window, providing structurally sound and informationally complete input data for subsequent nonlinear transformations and dimensionality optimization of features.

[0017] To enhance the nonlinear representation capability of LNN networks for leaf health status, the ReLU activation function is introduced to perform nonlinear transformation processing on the integrated features. The ReLU function suppresses ineffective negative feature responses, strengthens the expression weights of core effective features, and improves the learning and fitting efficiency of LNN networks for deep features, enabling them to accurately adapt to the complex nonlinear changes in leaf health status during operation.

[0018] Furthermore, the LightGBM model improved with particle swarm optimization (PSO) and cross-validation mechanisms, as described in S4, primarily relies on PSO to optimize key model parameters. This algorithm simulates the cooperative search behavior of a group of organisms, treating each parameter combination as a particle. Through information sharing and position updates among particles, it efficiently explores the parameter space comprised of the learning rate, tree depth, and number of leaf nodes. Its swarm intelligence characteristic avoids the problem of traditional parameter tuning getting stuck in local optima, quickly locating the global optimum and providing a more comprehensive parameter configuration foundation for the LightGBM model to improve its performance.

[0019] The LightGBM model is enhanced by incorporating a cross-validation strategy. By dividing the original dataset into multiple mutually exclusive subsets and alternating between these subsets as training and validation sets during training, the model can be adequately trained and validated under varying data distributions. This approach reduces the impact of data partitioning bias on model performance evaluation, improves the model's adaptability to complex variations in leaf data, and maintains reliable classification results.

[0020] Furthermore, as described in S4, the LightGBM model is used for deep classification learning of high-dimensional feature vectors. The model employs a histogram-based feature binning strategy to reduce computational complexity and memory consumption. Simultaneously, it focuses on key features and samples through gradient unilateral sampling and leaf growth strategies. The deep classification learning capability of the LightGBM model can fit feature patterns under different health states, achieving an effective mapping from the high-dimensional feature space to health state categories.

[0021] Furthermore, the blade health status levels described in S4 include four levels: healthy, slightly abnormal, moderately abnormal, and severely abnormal. In the healthy state, all monitoring indicators of the blade are stable within the normal range, with no obvious signs of damage. In the slightly abnormal state, there are slight characteristic deviations, possibly caused by minor wear or environmental interference, which do not currently affect operation. In the moderately abnormal state, the characteristic deviations are significant, and the blade shows localized minor damage, requiring enhanced monitoring. In the severely abnormal state, the characteristics deviate severely from the normal range, and the blade faces a significant risk of damage or failure, requiring immediate shutdown and maintenance.

[0022] The above description is merely an overview of the technical solution of this application. To better understand the technical means of this application and to facilitate its implementation according to the description, and to make the above and other objects, features, and advantages of this application more apparent, specific embodiments of this application are provided below. It should be understood that the above general description and the following detailed description are merely exemplary and do not limit the invention. Attached Figure Description

[0023] To more clearly illustrate the technical solutions implemented in this invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of this invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0024] Figure 1 This is a flowchart illustrating the operational process of a wind turbine blade health status assessment method based on LNN and an improved LightGBM, as described in this invention.

[0025] Figure 2 This is a flowchart illustrating the feature fusion of acoustic and vibration features in a wind turbine blade health status assessment method based on LNN and an improved LightGBM, as described in this invention.

[0026] Figure 3 This is a flowchart illustrating the operational process of leaf health status assessment using an LNN network and an improved LightGBM model as described in this invention. Detailed Implementation

[0027] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0028] This embodiment provides a method for assessing the health status of wind turbine blades based on LNN and an improved LightGBM, such as... Figure 1 As shown, the method includes the following steps: data acquisition and preprocessing, extraction of acoustic and vibration signal features, feature fusion using an attention mechanism, spatiotemporal alignment processing, construction of a multi-dimensional feature matrix, normalization processing, feature extraction via an LNN network, and deep classification learning using an improved LightGBM model to evaluate the health status level of wind turbine blades. The specific implementation plan is as follows: S1. Acoustic sensors deployed at the root of the wind turbine blades and three vibration signal sensors deployed at the junction of the hub and the blades are used to acquire acoustic and vibration signals for each blade, respectively. This data is then used to obtain SCADA data of the wind turbine's operating status within the wind turbine monitoring system. Preprocessing of the acoustic, vibration, and SCADA data is performed, specifically including wavelet thresholding and Fourier transform for the blade acoustic signals; low-pass filtering, linear interpolation, and detrending processing for the vibration signals; and trend separation, moving average filtering, outlier detection and correction, interpolation compensation, and normalization processing for the SCADA data.

[0029] S2, such as Figure 2 As shown, feature fusion is performed on the voiceprint signal and vibration signal to construct a multi-dimensional feature matrix with a unified time scale. First, feature extraction is performed on the preprocessed voiceprint signal and vibration signal in S1 to obtain voiceprint features and vibration features respectively. Both voiceprint features and vibration features include time-domain features and frequency-domain features: time-domain features include root mean square, variance, standard deviation, skewness, and kurtosis; frequency-domain features include dominant frequency amplitude, bandwidth, spectral entropy, spectral flatness, and spectral roll-off point.

[0030] Secondly, voiceprint features Vibration characteristics The input is fed into the attention mechanism module to calculate the correlation score between features: in, It is a simple, learnable function.

[0031] Next, the score is normalized using the Softmax function: This yields a set of data that can determine the differences in the contribution of acoustic and vibration signals to fault diagnosis or condition monitoring tasks under different operating conditions. and Different weights will be automatically assigned.

[0032] Finally, the extracted voiceprint features and vibration features are multiplied element-wise with their corresponding dynamic weights and then superimposed to obtain a complementary acoustic-vibration fusion feature vector.

[0033] Attention weights are multiplied element-wise with their corresponding features: The result of adding the two is This refers to the fused acoustic-vibration fusion feature vector: Next, using the sampling time point of the SCADA data as a reference, the acoustic-vibration fusion feature vector is spatiotemporally aligned to ensure that the acoustic signature signal, vibration signal, and SCADA data are consistent in time scale. Then, the acoustic-vibration fusion feature vector is associated and reconstructed with the preprocessed SCADA data to obtain a multidimensional feature matrix.

[0034] The multidimensional feature matrix is ​​normalized using the Z-score normalization method: in, These are the original data points. It is the mean of the sample data. It is the standard deviation of the sample data.

[0035] Z-score normalization can convert different features into a standard distribution with zero mean and unit variance, effectively eliminating the influence of dimensions between features.

[0036] S3, such as Figure 3 As shown, a high-dimensional feature vector representing the blade's health status is obtained by dynamically extracting deep features from the multidimensional feature matrix using an LNN network. Firstly, the core advantage of the LNN network stems from the accurate capture of the correlation between acoustic and vibration fusion data through the local receptive field mechanism. Under different blade health states, the changes in acoustic and vibration signals at adjacent time steps exhibit differentiated correlation strengths: in the early stages of a fault, signal fluctuations are smooth, with correlations between adjacent time steps exceeding 0.8; as the fault worsens, signal abrupt changes occur frequently, and significant feature differences emerge within local time windows.

[0037] Therefore, the receptive field size and step size need to be dynamically set based on the characteristics of the data time scale. This adaptive window strategy simulates a local perception mode, performing connection calculations only on features at adjacent time steps. Compared with fully connected networks, it reduces the number of parameters, avoids interference from irrelevant information caused by global connections, and can extract local dependencies at different time scales in a hierarchical manner, providing basic data for subsequent feature processing.

[0038] The fully connected layer acts as a feature integration hub, achieving systematic fusion of multi-dimensional features through structured information interaction. After the multi-dimensional feature matrix is ​​extracted by the local receptive field mechanism of the LNN network, it forms several independent local feature subsets, each corresponding to dimensional information within a specific time window. The fully connected layer constructs multiple learnable weight matrices to map these scattered feature subsets to a unified high-dimensional feature space. The fully connected layer includes a first fully connected layer, intermediate fully connected layers, and a final fully connected layer. Each local feature subset achieves feature dimension alignment after passing through the first fully connected layer, converting local features from different receptive fields into intermediate features of the same dimension. The intermediate features, after passing through the intermediate fully connected layer, achieve cross-time window information complementarity through matrix multiplication to obtain cross-window fused features. The cross-window fused features, after passing through the final fully connected layer, output a normalized integrated feature vector. Furthermore, to avoid overfitting during feature integration, a Dropout regularization mechanism is introduced: randomly discarding a portion of neuron connections forces the LNN network to learn more generalized feature association patterns. This integration mechanism not only fully preserves the temporal specificity of each local feature, but also effectively eliminates the information limitations of a single local window, enabling the output features to possess both temporal correlation and dimensional completeness, thus providing a stable input for subsequent nonlinear transformations.

[0039] To enhance the network's ability to nonlinearly represent the health status of leaves, the LNN network introduces... Activation functions and improved strategies are used to process integrated features: That is when The original value is retained when... The output is 0. The output obtained by using the ReLU activation function as a nonlinear transformation in an LNN neural network is: The mapping relationship between the health status of blade acoustic vibration signals exhibits strong nonlinearity: the signal characteristics of the same fault type differ significantly at different rotational speeds, while different fault types may show similar local signal patterns. The ReLU function, by setting the negative feature response to zero, can adaptively suppress noise interference and invalid features, and maximize the filtering ability of neurons. Through this nonlinear transformation, the deep feature learning efficiency of the LNN network is significantly improved, enabling it to adapt to the complex nonlinear changes in blade health status during operation, and ultimately outputting a high-dimensional feature vector representing the blade health status.

[0040] S4 uses the high-dimensional feature vectors obtained in S3 to perform deep classification learning on the LightGBM model, which is improved by particle swarm optimization (PSO) and cross-validation, to obtain the health status assessment level of wind turbine blades. PSO simulates the cooperative search behavior of a group of organisms, treating each parameter combination as a particle. Through information sharing and position updates among particles, it efficiently explores the parameter space composed of learning rate, tree depth, and the number of leaf nodes. Cross-validation enhances the LightGBM model by dividing the original dataset into multiple mutually exclusive subsets. During training, different subsets are alternately selected as training and validation sets, ensuring the model is adequately trained and validated under different data distributions.

[0041] The LightGBM model minimizes the weighted sum of empirical loss and regularization term by iteratively fitting the negative gradient: Let the model score be probability Then the cross-entropy loss is: The first-order gradient and the second-order gradient are: The LightGBM model employs a histogram-based feature binning strategy, discretizing persistent features into a finite number of B bins and accumulating gradients within each bin. This avoids the need for precise ranking of each sample. Arrange the samples according to their respective boxes. polymerization: When splitting a feature, only the bin-level values ​​need to be accumulated. H, It also calculates the split gain, no longer relying on the original sample-by-sample sorting.

[0042] Through gradient unilateral sampling and leaf growth strategies, the LightGBM model's deep classification learning capability can fit feature patterns under different health states, achieving an effective mapping from high-dimensional feature space to health state categories, and ultimately completing the assessment of the health status level of wind turbine blades.

[0043] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions 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 assessing the health status of wind turbine blades based on LNN and an improved LightGBM, characterized in that, Includes the following steps: S1: Acoustic and vibration signals of the blades are collected by acoustic and vibration sensors, respectively; the acoustic and vibration signals, as well as the SCADA data of the wind turbine operating status obtained from the wind turbine monitoring system, are preprocessed. S2: Extract features from the preprocessed voiceprint signal and vibration signal to obtain voiceprint features and vibration features respectively. Input the voiceprint features and vibration features into the attention mechanism module to obtain the sound and vibration fusion feature vector. Associate the sound and vibration fusion feature vector with the preprocessed SCADA data to reconstruct a multi-dimensional feature matrix and perform normalization processing. S3: Input the normalized multidimensional feature matrix into the LNN network, and obtain a high-dimensional feature vector representing the health status of the leaf by dynamically extracting features from the multidimensional feature matrix through the LNN network. S4: The improved LightGBM model is used for deep classification learning of high-dimensional feature vectors to determine the health status level of wind turbine blades. S3 includes: The normalized multidimensional feature matrix is ​​input into the LNN network; After the multidimensional feature matrix is ​​extracted by the local receptive field mechanism of the LNN network, it forms several independent local feature subsets, each of which corresponds to the dimensional information within a specific time window. Each local feature subset is mapped to a unified high-dimensional feature space through a fully connected layer of an LNN network to obtain an integrated feature vector. The fully connected layer includes a first fully connected layer, intermediate fully connected layers, and a last fully connected layer. After passing through the first fully connected layer, the feature dimensions of each local feature subset are aligned, and local features from different receptive fields are converted into intermediate features of the same dimension. The intermediate features are then processed through the intermediate fully connected layer to achieve cross-time window information complementarity through matrix multiplication to obtain cross-window fused features. The cross-window fused features are then processed through the last fully connected layer to output a normalized integrated feature vector. The integrated feature vector is nonlinearly transformed by the activation function of the LNN network to obtain a high-dimensional feature vector characterizing the health status of the leaf. In S4, the improvements to the LightGBM model include using particle swarm optimization to optimize key hyperparameters of LightGBM such as learning rate, number of trees, maximum tree depth, and number of leaf nodes, and using cross-validation to enhance the LightGBM model by dividing the original dataset into multiple mutually exclusive subsets and alternately selecting different subsets as training and validation sets during the training process.

2. The wind turbine blade health status assessment method based on LNN and improved LightGBM according to claim 1 is characterized as follows: the acoustic fingerprint sensor in S1 is deployed at the root of the three blades of the wind turbine, and the vibration sensor is deployed at the junction of the hub and the three blades and extends into the interior of the blade.

3. The wind turbine blade health status assessment method based on LNN and improved LightGBM according to claim 1 is characterized as follows: the SCADA data in S1 includes turbulence intensity, wind speed, blade rotation speed, hub vibration value, generator power, generator rotation speed, pitch shaft rate and pitch shaft position deviation.

4. The method for assessing the health status of wind turbine blades based on LNN and an improved LightGBM as described in claim 1. Its characteristics are as follows: In S1: Preprocessing the voiceprint signal includes: performing wavelet threshold denoising and Fourier transform on the voiceprint signal; Preprocessing the vibration signal includes: low-pass filtering, linear interpolation, and detrending processing of the vibration signal; Preprocessing of the SCADA data includes: trend separation, moving average filtering, outlier detection and correction, interpolation compensation and normalization.

5. The method for assessing the health status of wind turbine blades based on LNN and improved LightGBM as described in claim 1, characterized in that: the acoustic features and vibration features mentioned in S2 both include time-domain features and frequency-domain features; The time-domain features include root mean square, variance, standard deviation, skewness, and kurtosis; The frequency domain features include dominant frequency amplitude, bandwidth, spectral entropy, spectral flatness, and spectral roll-off point.

6. The wind turbine blade health status assessment method based on LNN and improved LightGBM according to claim 1, characterized in that: in S2, inputting the acoustic signature features and vibration features into the attention mechanism module to obtain the acoustic-vibration fusion feature vector includes: The voiceprint features and vibration features are input into the attention mechanism module to calculate the correlation scores between the voiceprint features and vibration features respectively: The correlation score was normalized using the Softmax function to obtain the dynamic weight of the voiceprint and the dynamic weight of the vibration. The voiceprint features and vibration features are multiplied element-wise with the dynamic weights of the voiceprint features and the dynamic weights of the vibration features, and then superimposed to obtain the sound-vibration fusion feature vector.

7. The wind turbine blade health status assessment method based on LNN and improved LightGBM according to claim 1, characterized in that: the multi-dimensional feature matrix obtained by associating and reconstructing the acoustic-vibration fusion feature vector with the preprocessed SCADA data in S2 includes: Based on the sampling time point of SCADA data, the acoustic-vibration fusion feature vector is spatiotemporally aligned, and then the acoustic-vibration fusion feature vector and SCADA data are used to construct a multi-dimensional feature matrix to realize the correlation reconstruction of SCADA data with acoustic features and vibration features.

8. The wind turbine blade health status assessment method based on LNN and improved LightGBM according to claim 1 is characterized as follows: In S2, the multidimensional feature matrix is ​​normalized using the Z-score normalization method.

9. The wind turbine blade health status assessment method based on LNN and improved LightGBM according to claim 1, characterized as follows: The wind turbine blade health status levels in S4 include four levels: healthy state, slightly abnormal state, moderately abnormal state, and severely abnormal state; healthy state indicates that all monitoring indicators of the blade are stable within the normal range and there are no obvious signs of damage; slightly abnormal state indicates that there is a slight characteristic deviation, which may be caused by minor wear or environmental interference and does not affect operation for the time being; moderately abnormal state indicates that the characteristic deviation is significant and the blade has local minor damage, requiring enhanced monitoring; severely abnormal state indicates that the characteristics deviate significantly from the normal range, the blade has a greater risk of damage or failure, and immediate shutdown and maintenance are required.

Images