Wind turbine blade efficiency diagnosis method and device based on field operation data

Abstract

This invention provides a method and apparatus for diagnosing the performance of wind turbine blades based on field operation data. It relates to the technical field of wind turbine blade performance diagnosis methods, including feature extraction from strain time-series data to obtain temporal and spatial vectors; calculating first, second, and third weight coefficients using a gated attention weight dynamic allocation method to obtain a fused vector of the blade in its current state; calculating the health entropy value of the blade at each moment using the information entropy method, setting a threshold range for the entropy value, and classifying the blade's health entropy values. The extracted temporal vector captures the gradual characteristics of structural fatigue, and the extracted spatial vector represents the local damage patterns of the acoustic spectrum. The weights are dynamically allocated through a gated attention mechanism to avoid misjudgment based on a single signal, and the blade damage level is dynamically distinguished using the health entropy value.

Description

Technical Field

[0001] This invention relates to the technical field of wind turbine blade performance diagnosis methods, specifically to a wind turbine blade performance diagnosis method and device based on field operation data. Background Technology

[0002] The current wind turbine blade health monitoring technology relies solely on strain or vibration single-mode data, which is insufficiently sensitive to early micro-cracks and makes it difficult to distinguish between aerodynamic noise and actual damage signals. The use of fixed weight coefficients to fuse multi-source data is prone to feature aliasing under dynamic operating conditions such as pitch and yaw, leading to delayed warnings. Health assessment methods based on simple thresholds or normal distribution assumptions are difficult to adapt to non-Gaussian distributed data in complex environments such as marine salt spray corrosion and freezing.

[0003] In the prior art, patent document CN112343776A discloses a method for verifying performance using existing SCADA data from the field and constructing a transfer function based on blade pitch angle and blade performance. However, this method does not integrate strain time-series data and acoustic pressure Mel spectrum features to improve detection sensitivity, does not use a gated attention weight dynamic allocation mechanism to achieve adaptive optimization of feature vectors, does not use a weighted model of angle of attack deviation based on real-time calculation of axial wind speed to improve the aerodynamic load interference suppression rate, and does not use Gaussian kernel density estimation and information entropy classification to reduce the fitting error of non-Gaussian distributed data and support quantitative prediction of crack propagation rate. Therefore, there is an urgent need for a wind turbine blade performance diagnosis method and device based on field operation data.

[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide a method and apparatus for diagnosing the performance of wind turbine blades based on field operation data, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] The method for diagnosing wind turbine blade performance based on field operation data includes the following steps:

[0008] S1: Real-time acquisition of strain time series data and sound pressure data of each blade of the wind turbine to be diagnosed, Fourier transform of the sound pressure data to obtain the linear spectrum and determine the power spectrum corresponding to the linear spectrum, compress the power spectrum to generate a grayscale image, synchronously acquire the radial wind speed of the blade, calculate the angle of attack at different positions of the blade and use the weighted average method to obtain the overall angle of attack deviation of the blade.

[0009] S2: Feature extraction is performed on the strain time series data to obtain time series vector and spatial vector. The gating attention weight dynamic allocation method is used to dynamically allocate the weights of the time series vector, spatial vector and angle of attack deviation. Based on the set weight coefficients, the time series vector, spatial vector and angle of attack deviation are concatenated to obtain the fusion vector of the current state of the blade.

[0010] S4: Based on the fusion vector of healthy wind turbine blade samples as the health reference vector, the fusion vector of the current state of each blade and the health reference vector are compared using the Mahalanobis distance method to obtain the deviation coefficient of the blade. Taking the current state time as the starting point, a time window and sampling frequency are set, and the deviation coefficient is sampled within the time window. The kernel density of each sampling point is calculated using a Gaussian kernel.

[0011] S5: Based on the kernel density of each sampling point, the health entropy value of the leaf within each time window is calculated using the information entropy method, and a threshold range for the entropy value is set to classify the health entropy value of the leaf.

[0012] Further, a grayscale image is generated, and the specific steps are as follows:

[0013] Sound pressure data is collected from the transition region of the blade from the root to the airfoil section. The sound pressure data is sampled at a preset sampling rate and processed in frames. Fourier transform is performed on each frame to obtain the linear amplitude spectrum of each frame. The power spectrum is obtained by squaring the linear amplitude spectrum. The energy of each frequency point in the power spectrum is multiplied by the response value of a Mel filter of preset dimension. The frequency range of the Mel filter covers 0 to half of the preset sampling rate. The results of each filter are then summed to obtain the power spectrum of the preset dimension Mel band. The power spectrum of each Mel band is compressed by taking the natural logarithm and normalized by Z-score. The normalized power spectrum is linearly mapped to the grayscale range of 0 to 255. A time window containing a preset number of frames is set and updated at a preset time frequency. A column of grayscale values ​​of preset dimension is generated for each frame. The time axis is stitched together along the horizontal direction to finally form a grayscale image with the preset number of frames as rows and the preset dimension as columns.

[0014] Furthermore, the weighted average method is used to calculate the overall angle of attack deviation of the blades. The specific steps are as follows:

[0015] The radial wind speed of the blades is acquired in real time, and the angle of attack deviation is calculated using the following formula:

[0016]

[0017] Indicates the axial wind speed of the blades; Indicates the blade rotational speed; Indicates the radius of rotation of the blade; Indicates the pitch angle; The radius of rotation of the blade is indicated by Angle of attack deviation;

[0018] The angle of attack at 25%, 50%, and 75% of the blade radius is calculated in real time, and a weighted average is calculated as the overall angle of attack deviation of the blade.

[0019]

[0020] in, This indicates the real-time overall angle of attack deviation of the blade.

[0021] Further, the time-series vector is extracted, and the specific steps are as follows:

[0022] The location of the main beam axis at which the historical blade is obtained is at a distance from the hub. Strain time-series data samples were collected at the location, and the health status of the blades was classified by category labeling. Undamaged blades were marked as normal, and blades with crack lengths ranging from... The blades were marked as microcracks, and the crack length was greater than... Leaves are marked as severely damaged; the category labels are converted to One-Hot encoding, specifically: normal leaves are marked as... Microcracks are Severe damage is ,in Indicates the total length of the leaf;

[0023] Set a sliding time window and a number of time steps, moving at fixed time intervals. Add a classification head to the data, using the strain time series data samples within the sliding time window as... The input will The predicted vector output from the hidden layer is mapped to the corresponding label probability. The label probability is compared with the corresponding One-Hot encoding, the cross-entropy loss is calculated, and the loss is adjusted inversely. The training is completed when the loss entropy is less than a preset value, along with the parameters of the classification head.

[0024] Training completed Remove the classification head in the middle, and position the main beam axis of the blade at a distance from the hub. Real-time strain time-series data of the location is input into the trained system. In the network, extract the time-series vector of the last time step within the window.

[0025] Furthermore, the fusion vector of the blade in the current state is obtained, specifically as follows:

[0026] Input the grayscale image into the process of removing fully connected layers. In the network, the spatial vector is obtained. Based on the temporal vector and the spatial vector, the first, second, and third weight coefficients are calculated using a gated attention weight dynamic allocation method, as follows:

[0027] The temporal and spatial vectors are linearly projected through different weight matrices, and the projection results are summed to obtain a joint feature vector. A hyperbolic tangent function is then applied to the joint feature vector for activation, mapping the numerical range to... In the interval, the three preset attention weight vectors are used to perform dot product operations with the activated joint feature vector, and the result is then processed. The values ​​of the obtained vector elements are used as the first, second, and third weighting coefficients, respectively;

[0028] The vector concatenation method is used to concatenate the time-series vector, spatial vector, and angle-of-attack deviation. Specifically, the first, second, and third weight coefficients are multiplied by the time-series vector, spatial vector, and angle-of-attack deviation, respectively. The product is then expanded into an equal-dimensional vector by padding, i.e., all padding values ​​are set to 0. The padded vectors are then concatenated sequentially by connecting the first and last ends to obtain the fusion vector of the blade in the current state.

[0029] Furthermore, the deviation coefficient of the blade in the current state is calculated using the Mahalanobis distance method. The specific steps are as follows:

[0030] Obtain the fusion vector of historical healthy wind turbine blade samples, calculate the mean vector of all samples as the health baseline vector, and calculate the deviation coefficient of the current state blades using the Mahalanobis distance method based on the health baseline vector and the fusion vector.

[0031] Subtract the current state's fusion vector from the health baseline vector dimension by dimension to obtain the deviation for each dimension. Based on the deviation, calculate the covariance matrix of the fusion vector of the historical healthy wind turbine blade samples. Invert the covariance matrix and use the inverted matrix as the weight matrix. Weight the deviation for each dimension and take the square root of the weighted result to obtain the deviation coefficient.

[0032] Furthermore, the kernel density of each sampling point is calculated using a Gaussian kernel. The specific steps are as follows:

[0033] Starting from the current state, a time window and sampling frequency are set. Within the time window, the deviation coefficients are sampled, and the kernel density of each sampling point is calculated using a Gaussian kernel.

[0034]

[0035] in, Indicates the time window, the first Kernel density of each sampling point; Indicates the first Deviation coefficient for each sampling point; Indicates the time window, the first Deviation coefficient for each sampling point; Indicates the number of sampling points; Indicates Gaussian bandwidth;

[0036] Based on the kernel density of each sampling point, the health entropy value within the leaf time window is calculated using the information entropy method, and a threshold range for the entropy value is set to classify the health entropy value:

[0037]

[0038] in, This represents the health entropy value within the leaf's time window;

[0039] when A value greater than 3 and less than or equal to 4.5 indicates that the blade has a micro-crack. A value greater than 4.5 indicates severe structural damage to the blade. A value of 3 or less indicates that the leaf blades are normal.

[0040] The present invention also provides a wind turbine blade performance diagnostic device based on field operation data. The diagnostic device is used to perform the above-described diagnostic method, including:

[0041] The data acquisition module is used to collect the strain time series data and sound pressure data of each blade of the wind turbine to be diagnosed in real time. It performs Fourier transform on the sound pressure data to obtain the linear spectrum and determines the power spectrum corresponding to the linear spectrum. Based on the power spectrum, it compresses to generate a grayscale image, simultaneously acquires the radial wind speed of the blade, calculates the angle of attack at different positions of the blade, and uses the weighted average method to obtain the overall angle of attack deviation of the blade.

[0042] The fusion vector acquisition module is used to extract features from strain time series data to obtain time series vectors and spatial vectors. It uses a gated attention weight dynamic allocation method to dynamically allocate the weights of time series vectors, spatial vectors and angle of attack deviations. Based on the set weight coefficients, the time series vectors, spatial vectors and angle of attack deviations are concatenated to obtain the fusion vector of the current state of the blade.

[0043] The deviation coefficient calculation module is used to compare the fusion vector of the healthy wind turbine blade sample with the healthy reference vector using Mahalanobis distance to obtain the deviation coefficient of the blade. Taking the current state time as the starting point, a time window and sampling frequency are set, and the deviation coefficient is sampled within the time window. The kernel density of each sampling point is calculated using a Gaussian kernel.

[0044] The comparison and classification module is used to calculate the health entropy value of the leaf within each time window based on the kernel density of each sampling point using the information entropy method, and to set the threshold range of the entropy value to classify the health entropy value of the leaf.

[0045] Compared with the prior art, the beneficial effects of the present invention are:

[0046] The detection sensitivity of cracks is improved by fusing strain time-series data with acoustic pressure Mel spectrum features; the gating attention weight dynamic allocation mechanism enables adaptive optimization of feature vectors; the angle-of-attack deviation weighted model based on real-time axial wind speed calculation improves the aerodynamic load interference suppression rate; the fitting error of non-Gaussian distributed data is reduced by using Gaussian kernel density estimation and information entropy classification, supporting quantitative prediction of crack propagation rate; the grayscale image compression algorithm reduces the amount of data transmission, and the lightweight LSTM-CNN model improves the inference speed. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the overall method flow of the present invention;

[0048] Figure 2 The graph shows the relationship between the angle of attack deviation and the axial wind speed.

[0049] Figure 3 This is a schematic diagram of the overall device of the present invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0051] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0052] Example:

[0053] Please see Figures 1-2 The present invention provides a technical solution:

[0054] The method for diagnosing wind turbine blade performance based on field operation data includes the following steps:

[0055] S1: Real-time acquisition of strain time series data and sound pressure data of each blade of the wind turbine to be diagnosed, Fourier transform of the sound pressure data to obtain the linear spectrum and determine the power spectrum corresponding to the linear spectrum, compress the power spectrum to generate a grayscale image, synchronously acquire the radial wind speed of the blade, calculate the angle of attack at different positions of the blade and use the weighted average method to obtain the overall angle of attack deviation of the blade.

[0056] The specific steps to generate a grayscale image are as follows:

[0057] Sound pressure data is collected from the transition region of the blade from the root to the airfoil section. The sound pressure data is sampled at a preset sampling rate and processed in frames. Fourier transform is performed on each frame to obtain the linear amplitude spectrum of each frame. The power spectrum is obtained by squaring the linear amplitude spectrum. The energy of each frequency point in the power spectrum is multiplied by the response value of a Mel filter of preset dimension. The frequency range of the Mel filter covers 0 to half of the preset sampling rate. The results of each filter are then summed to obtain the power spectrum of the preset dimension Mel band. The power spectrum of each Mel band is compressed by taking the natural logarithm and normalized by Z-score. The normalized power spectrum is linearly mapped to the grayscale range of 0 to 255. A time window containing a preset number of frames is set and updated at a preset time frequency. A column of grayscale values ​​of preset dimension is generated for each frame. The time axis is stitched together along the horizontal direction to finally form a grayscale image with the preset number of frames as rows and the preset dimension as columns.

[0058] The core technical effect achieved in the above process is as follows: the high-dimensional, non-stationary one-dimensional sound pressure time-series signal is transformed into a two-dimensional grayscale image rich in time-frequency texture features. This transformation not only simulates the nonlinear auditory perception characteristics of the human ear through the Mel filter bank, mapping the blade aerodynamic noise and high-frequency howling of cracks, which are concentrated in the low-frequency region, to a more sensitive frequency band, thus achieving targeted feature enhancement; at the same time, by taking the logarithmic compression of the power spectrum and Z-score normalization, the interference caused by the difference in the absolute amplitude of the signal under different wind speeds and operating conditions is greatly suppressed, and the relative spectral changes reflecting the blade structural state are amplified; the resulting grayscale image with time as the horizontal axis and the Mel frequency band as the vertical axis is essentially a "time-frequency heatmap" of the blade's acoustic fingerprint, providing high-quality input for the subsequent automatic extraction of deep spatial features by the convolutional neural network (CNN).

[0059] Mel-frequency spectral compression and grayscale mapping are essentially robust nonlinear normalization techniques: logarithmic compression maps the power spectrum with a wide dynamic range to the level of "loudness" perceived by the human ear, while Z-score normalization eliminates the drift of the signal energy baseline across different time periods, allowing CNNs to focus on the energy distribution patterns between frequency bands rather than absolute intensity. Furthermore, stitching multiple consecutive frames into a grayscale image introduces temporal contextual information, enabling transient sounds such as cracks opening and closing at specific moments of stress to leave coherent texture traces across several frames, significantly improving the detection probability of weak fault signals.

[0060] The weighted average method is used to calculate the overall angle of attack deviation of the blades. The specific steps are as follows:

[0061] The radial wind speed of the blades is acquired in real time, and the angle of attack deviation is calculated using the following formula:

[0062]

[0063] Indicates the axial wind speed of the blades; Indicates the blade rotational speed; Indicates the radius of rotation of the blade; Indicates the pitch angle; The radius of rotation of the blade is indicated by Angle of attack deviation;

[0064] The above formula establishes the mapping relationship between various operating parameters and the angle of attack based on aerodynamic principles. Among them, The actual inflow angle characterizing the airflow relative to the blade motion: In other words, the axial wind speed determines the velocity component of the airflow impacting the blades along the wind turbine axis. That is, the tangential linear velocity determines the tangential velocity component of the airflow cut by the rotating blade, and the ratio of the two, after arctangent calculation, is physically the angle between the incoming direction of the airflow and the plane of rotation. The pitch angle, representing the twist angle of the blade itself, is the geometric angle between the airfoil chord and the plane of rotation, and the actual inflow angle. Subtracting the pitch angle, the difference is the true angle of attack of the airflow relative to the airfoil chord; axial wind speed and There is a positive correlation: when the rotational speed and blade angle remain constant, an increase in axial wind speed increases the inflow angle, and the angle of attack deviation increases accordingly; tangential velocity... and There is a negative correlation: when wind speed and propeller angle remain constant, increasing the rotational speed increases the tangential velocity, decreases the inflow angle, and consequently reduces the angle of attack deviation; pitch angle adjustment and Negative correlation: When wind speed and rotational speed remain constant, increasing the propeller angle increases the airfoil geometry angle, directly reducing the true angle of attack, and consequently reducing the angle of attack deviation.

[0065] In the above embodiments, 20 sets of data on axial wind speed and corresponding angle of attack deviation are given to reflect the change of angle of attack deviation with axial wind speed, as shown in Table 1:

[0066] Table 1: Relationship between Angle of Attack Deviation and Axial Wind Speed

[0067]

[0068] In Table 1 above, given , , Under these conditions, axial wind speed is positively correlated with angle of attack deviation. When the rotational speed and propeller angle remain constant, an increase in axial wind speed increases the inlet angle, and the angle of attack deviation increases accordingly.

[0069] The angle of attack at 25%, 50%, and 75% of the blade radius is calculated in real time, and a weighted average is calculated as the overall angle of attack deviation of the blade.

[0070]

[0071] in, This indicates the real-time overall angle of attack deviation of the blade.

[0072] In the above process, by weighted fusion of the local angle-of-attack deviations of three key airfoil sections—blade root, blade mid-section, and blade tip—a comprehensive index capable of characterizing the overall aerodynamic load state of the blade is constructed. The reason for adopting , , The asymmetric distribution coefficient is based on a comprehensive consideration of the aerodynamic load distribution pattern during blade operation and the structural damage-sensitive area: The highest weight is assigned at the radius. This is because the region is located in the middle section of the blade, which avoids the interference of complex flow at the blade root and vortex at the blade tip. It is also the core area where the blade bears the maximum bending moment and captures the main energy. Its angle of attack deviation has the most significant impact on aerodynamic efficiency. Assigned at the radius Although it is close to the blade root, the airfoil in this area is thick, relatively insensitive to changes in angle of attack, and is greatly affected by the flow around the hub. Assigned at the radius Located near the blade tip, the blade has a high linear velocity and is sensitive to aerodynamic noise. However, the presence of tip vortices makes angle of attack measurements susceptible to turbulent disturbances. This weighted strategy achieves an optimized balance in terms of physical mechanism, focusing on the core energy production region, taking into account the safe region of the blade root structure, and appropriately introducing the tip sensitive region. This ensures that the overall angle of attack deviation after fusion can sensitively reflect the deterioration of blade aerodynamic performance without causing drastic fluctuations due to local instantaneous turbulence.

[0073] S2: Feature extraction is performed on the strain time series data to obtain time series vector and spatial vector. The gating attention weight dynamic allocation method is used to dynamically allocate the weights of the time series vector, spatial vector and angle of attack deviation. Based on the set weight coefficients, the time series vector, spatial vector and angle of attack deviation are concatenated to obtain the fusion vector of the current state of the blade.

[0074] The specific steps for extracting time series vectors are as follows:

[0075] The location of the main beam axis at which the historical blade is obtained is at a distance from the hub. Strain time-series data samples were collected at the location, and the health status of the blades was classified by category labeling. Undamaged blades were marked as normal, and blades with crack lengths ranging from... The blades were marked as microcracks, and the crack length was greater than... Leaves are marked as severely damaged; the category labels are converted to One-Hot encoding, specifically: normal leaves are marked as... Microcracks are Severe damage is ,in Indicates the total length of the leaf;

[0076] Set a sliding time window and a number of time steps, moving at fixed time intervals. Add a classification head to the data, using the strain time series data samples within the sliding time window as... The input will The predicted vector output from the hidden layer is mapped to the corresponding label probability. The label probability is compared with the corresponding One-Hot encoding, the cross-entropy loss is calculated, and the loss is adjusted inversely. The training is completed when the loss entropy is less than a preset value, along with the parameters of the classification head.

[0077] Training completed Remove the classification head in the middle, and position the main beam axis of the blade at a distance from the hub. Real-time strain time-series data of the location is input into the trained system. In the network, extract the time-series vector of the last time step within the window.

[0078] In the above process, the LSTM network is forced to learn to extract deep temporal features highly correlated with the degree of blade damage from the original strain time series data through pre-training classification tasks, thereby obtaining a "feature extractor" with physical semantics. The reason for selecting a position 0.3L from the hub at the main beam axis to collect strain data is that this section is located at the intersection of the maximum chord length region of the blade and the aerodynamic center. It is a key structural part with the largest bending moment in the blade flapping direction and the most concentrated fatigue damage accumulation. Placing sensors at this location can obtain strain response signals that are most sensitive to crack initiation and propagation. The healthy state is divided into three levels: normal, microcracks, and severe damage. One-Hot encoding is used as the supervision signal. The technical purpose is to construct a progressive damage classification task with clear engineering boundaries. That is, the 0.5mm threshold corresponds to the visual detection limit and the critical point of coating cracking, and the 2mm threshold corresponds to the engineering experience limit of structural adhesive layer failure and crack entering the unstable propagation stage. Through training on this classification task, the LSTM network is forced to learn to distinguish the temporal waveform differences of three essentially different states: "no damage", "repairable damage" and "irreversible damage". This makes the temporal vectors extracted after removing the classification head naturally have the ability to identify the degree of damage: the strain temporal vectors of normal blades are clustered in a specific region of the feature space, the vectors of small cracks show a distribution that is deviated but still close, and the vectors of severe damage are significantly outside the healthy clusters, thus providing feature inputs with clear physical discrimination for subsequent Mahalanobis distance comparison.

[0079] Suppose a batch contains The nth sample, the th The logits vector output by the classification head after each sample passes through the LSTM hidden layer is: The predicted probability after Softmax mapping is:

[0080]

[0081] Indicates the index number of the sample within the batch; Indicates the first The logits vector output by the classification head after each sample passes through the LSTM hidden layer, where , , The original scores correspond to the three categories of "normal", "minor cracks" and "severe damage" respectively;

[0082] The true One-Hot encoding of this sample is: The batch average cross-entropy loss is:

[0083]

[0084] in, This represents the batch average cross-entropy loss; Indicates the first After Softmax mapping, the sample belongs to the th sample. Predicted probability of class ; Indicates the first The true label One-Hot encoded vectors of each sample, where and ;

[0085] make Represents all trainable parameters of the LSTM network and the classification head, and the loss. For any gradient The calculation is performed by backpropagation along time, and the parameters are updated using the Adam optimizer:

[0086]

[0087]

[0088]

[0089]

[0090]

[0091] in, Indicates the learning rate; , These are the exponential decay rates of first-order momentum and second-order momentum, respectively. Represents the control constant; Indicates the first First-order momentum estimation of the step; Indicates the first First-order momentum estimation of the step; Indicates the first Second-order momentum estimation of the step; Indicates the first Second-order momentum estimation of the step; Indicates the first First-order momentum estimation after step bias correction; Indicates the first Second-order momentum estimation after step bias correction; express of The power is used for deviation correction; express of The power is used for deviation correction; Indicates the current iteration step;

[0092] Let the preset loss threshold be Training will terminate when the following condition is met: .

[0093] Removing the classification head from the trained LSTM achieves the core technical effect of reconstructing the LSTM network from a "damage classifier" to a "deep feature extractor," thus shifting from a "discrimination task" to a "representation task." During pre-training, the classification head and cross-entropy loss work together to force the LSTM's hidden layers to learn discriminative temporal features that can distinguish between normal, micro-cracks, and severe damage states. This process essentially involves supervised pre-training of the network parameters to enhance their damage sensitivity. Removing the classification head during actual diagnosis is necessary because the true damage labels of real-time data are unknown. Continuing to use the classification head would force the network to rigidly categorize the current state into three preset categories, failing to identify ambiguous states within threshold boundaries or quantify the degree of deviation from the health baseline. By removing the classification head, the LSTM ontology, pre-trained with damage semantics, is retained. Its output temporal vector no longer represents the "probability of belonging to a certain category" but rather a high-dimensional feature representation containing all sensitive information related to damage evolution in the strain time-series waveform.

[0094] To obtain the fusion vector of the blade in the current state, specifically:

[0095] Input the grayscale image into the process of removing fully connected layers. In the network, the spatial vector is obtained. Based on the temporal vector and the spatial vector, the first, second, and third weight coefficients are calculated using a gated attention weight dynamic allocation method, as follows:

[0096] The temporal and spatial vectors are linearly projected through different weight matrices, and the projection results are summed to obtain a joint feature vector. A hyperbolic tangent function is then applied to the joint feature vector for activation, mapping the numerical range to... In the interval, the three preset attention weight vectors are used to perform dot product operations with the activated joint feature vector, and the result is then processed. The values ​​of the obtained vector elements are used as the first, second, and third weighting coefficients, respectively;

[0097] The vector concatenation method is used to concatenate the time-series vector, spatial vector, and angle-of-attack deviation. Specifically, the first, second, and third weight coefficients are multiplied by the time-series vector, spatial vector, and angle-of-attack deviation, respectively. The product is then expanded into an equal-dimensional vector by padding, i.e., all padding values ​​are set to 0. The padded vectors are then concatenated sequentially by connecting the first and last ends to obtain the fusion vector of the blade in the current state.

[0098] The core technical effect of the above process is to achieve adaptive nonlinear fusion of multimodal features through a gated attention mechanism, ensuring that the fused vector always retains the feature modes most sensitive to damage under the current working condition and eliminates splicing distortion caused by inconsistent feature dimensions. Specifically, the temporal vector and spatial vector are linearly projected and added to construct a joint feature vector that fuses two types of deep features. After hyperbolic tangent activation, the three learnable attention weight vectors are multiplied and normalized respectively, thereby dynamically outputting the first, second, and third weight coefficients. These three coefficients are not fixed by humans, but are calculated in real time based on the "modal importance score" of the current input sample: when the strain time-series waveform shows obvious crack opening and closing characteristics, the first weight automatically increases; when the sound pressure Mel spectrum shows abnormal frequency band texture, the second weight dominates; when the angle of attack deviation deviates significantly due to pitch failure, the third weight is activated. Based on this, the weighted temporal vector, spatial vector and angle of attack deviation are expanded into equal-dimensional vectors by zero padding and then concatenated. This preserves the independent physical semantics of each modal feature and avoids the distortion of the feature space after concatenation caused by the original dimensional differences. The final fusion vector is a comprehensive representation of the blade health status that is adaptive to operating conditions, sensitive to damage, and dimensionally balanced.

[0099] The reason for inputting the grayscale image into a CNN network with the fully connected layers removed to obtain the spatial vector is that the essential function of the convolutional and pooling layers is local perception and spatial feature abstraction, rather than classification decision. That is, the grayscale image forms a two-dimensional time-frequency texture with time as the horizontal axis and Mel frequency band as the vertical axis. The CNN extracts the energy distribution pattern between different frequency bands by sliding the convolutional kernel in the frequency domain dimension and captures the evolution law of acoustic texture by sliding in the time dimension. After multiple layers of convolution and pooling, the original image is compressed into a feature map containing high-order time-frequency semantics. Flattening it gives the spatial vector representing the acoustic fingerprint of the leaf.

[0100] The determination mechanism for different weight matrices is as follows: the weight matrix used for linear projection and the three attention weight vectors are initialized as random vectors following a Gaussian distribution; during training, the Mahalanobis distance between the fusion vector of historical samples and the healthy baseline vector is used as the loss function, with the optimization objective being that the deviation of healthy samples approaches zero and the deviation of damaged samples increases significantly. The gradient of the loss function with respect to the parameters of each weight matrix is ​​calculated, where the weight matrix parameters include the weight matrix of linear projection and the three attention weight vectors. The weight matrix is ​​iteratively updated along the gradient descent direction. After training driven by massive amounts of data for dozens of iteration cycles, when the decrease in the loss function is less than the set tolerance in multiple consecutive cycles, it is determined that the weight matrix gradually converges to the optimal solution.

[0101] S4: Based on the fusion vector of healthy wind turbine blade samples as the health reference vector, the fusion vector of the current state of each blade and the health reference vector are compared using the Mahalanobis distance method to obtain the deviation coefficient of the blade. Taking the current state time as the starting point, a time window and sampling frequency are set, and the deviation coefficient is sampled within the time window. The kernel density of each sampling point is calculated using a Gaussian kernel.

[0102] S5: Based on the kernel density of each sampling point, the health entropy value of the leaf within each time window is calculated using the information entropy method, and a threshold range for the entropy value is set to classify the health entropy value of the leaf.

[0103] The deviation coefficient of the blade in its current state is calculated using the Mahalanobis distance method. The specific steps are as follows:

[0104] Obtain the fusion vector of historical healthy wind turbine blade samples, calculate the mean vector of all samples as the health baseline vector, and calculate the deviation coefficient of the current state blades using the Mahalanobis distance method based on the health baseline vector and the fusion vector.

[0105] Subtract the current state's fusion vector from the health baseline vector dimension by dimension to obtain the deviation for each dimension. Based on the deviation, calculate the covariance matrix of the fusion vector of the historical healthy wind turbine blade samples. Invert the covariance matrix and use the inverted matrix as the weight matrix. Weight the deviation for each dimension and take the square root of the weighted result to obtain the deviation coefficient.

[0106] In the above process, Mahalanobis distance is used to replace Euclidean distance to decouple and remove redundancy of the correlation between the dimensions of the fusion vector, thereby constructing a health deviation metric that is extremely sensitive to early minor damage to the blade and highly robust to fluctuations in operating conditions.

[0107] Specifically, there is a strong natural correlation between different monitoring feature dimensions of the blade. For example, an increase in the strain of the main beam is often accompanied by an increase in sound pressure in a specific frequency band. If the Euclidean distance is directly calculated, the repetitive information carried by these correlated dimensions will be repeatedly accumulated. At the same time, it is impossible to suppress the multi-dimensional synchronous drift caused by common reasons such as gusts and power grid fluctuations, resulting in an inflated deviation coefficient or masking the true damage. This step constructs a covariance matrix through the fusion vector of historical healthy samples. The physical essence of its inverse matrix as a weight matrix is ​​to apply a very small weight to the multi-dimensional synchronous offset that is consistent with the statistical distribution of historical healthy data in the direction of fluctuation, and to impose a great penalty on the abnormal offset that violates the historical correlation pattern. When a tiny crack appears on the blade, the correlation between strain and sound pressure will break the stable pattern in the healthy state. This "correlation structure destruction" is presented as a non-cooperative change in a specific direction in the deviation vector. The Mahalanobis distance, through the weighted transformation of the covariance inverse matrix, rotates and compresses the originally correlated feature space into an isotropic standard space, which significantly amplifies this weak non-cooperative anomaly. The deviation coefficient calculated in this way is no longer just "how much the amplitude deviates", but "to what extent the current state does not conform to the coordinated fluctuation law between the various characteristic dimensions of a healthy leaf", which greatly improves the ability to detect micro-cracks at an early stage under strong noise and strong interference.

[0108] The kernel density of each sampling point is calculated using a Gaussian kernel. The specific steps are as follows:

[0109] Starting from the current state, a time window and sampling frequency are set. Within the time window, the deviation coefficients are sampled, and the kernel density of each sampling point is calculated using a Gaussian kernel.

[0110]

[0111] in, Indicates the time window, the first Kernel density of each sampling point; Indicates the first Deviation coefficient for each sampling point; Indicates the time window, the first Deviation coefficient for each sampling point; Indicates the number of sampling points; Indicates Gaussian bandwidth;

[0112] In the above process, discrete deviation coefficient sampling points are reconstructed into a continuous probability density distribution through Gaussian kernel density estimation, thereby transforming the "instantaneous deviation amplitude" of the leaf health status into a quantitative characterization of the "time domain distribution pattern".

[0113] Specifically, observing only the deviation coefficient at a single moment can easily lead to false alarms due to sudden increases caused by non-fault factors such as gusts of wind or transient fluctuations in the power grid. This step, however, sets a time window starting from the current moment and performs kernel density estimation on the deviation coefficient sequence within the window. The technical essence is to assign a normally distributed "influence band" centered on itself to each sampling point using a Gaussian kernel function. By superimposing the influence bands of all sampling points, the true probability density curve of the deviation coefficient within that time window is fitted. (Gaussian bandwidth) In this process, the scale smoothing factor, i.e., the appropriate bandwidth, smooths the fluctuations in the deviation coefficient caused by random noise in healthy blades into a uniform and gentle density distribution, while the intermittent spikes in the deviation coefficient of damaged blades caused by crack opening and closing and aeroelastic instability are preserved as significant bulges on the density curve. The kernel density of each sampling point is calculated accordingly. Instead of representing the absolute magnitude of the deviation, it characterizes the relative frequency and clustering of the deviation within the window: the deviation coefficient of healthy blades, while fluctuating slightly, is evenly distributed with a gentle peak; the deviation coefficient of damaged blades, however, exhibits a clear concentrated distribution, with specific deviation amplitudes appearing repeatedly. This transformation provides a crucial probabilistic input for subsequent calculations of health entropy values, enabling the diagnostic model to leap from the amplitude dimension of "how much the blade deviated" to the statistical dimension of "what distribution pattern the blade's deviation behavior exhibits," significantly improving diagnostic stability under complex turbulent conditions.

[0114] Based on the kernel density of each sampling point, the health entropy value within the leaf time window is calculated using the information entropy method, and a threshold range for the entropy value is set to classify the health entropy value:

[0115]

[0116] in, This represents the health entropy value within the leaf's time window;

[0117] The technical effect of the above process is to further compress the probability distribution obtained from kernel density estimation into a single scalar—the health entropy value—thereby quantifying the "deviation coefficient fluctuation pattern" of the blade within the time window into a health level index with a physical threshold. Specifically, in information theory, entropy represents the degree of disorder or uncertainty of a system. The essential innovation of introducing this step into structural health monitoring lies in the fact that, under stable operating conditions, the deviation coefficient of a healthy blade should fluctuate slightly, uniformly, and randomly around the benchmark value. At this time, the kernel density distribution is flat, the probability of each sampling point is close, and the calculated BHE value is high, reflecting that the system is in a normal random fluctuation state of "high uncertainty and low regularity". However, when the blade has micro-cracks or severe damage, the deviation coefficient will exhibit a deterministic pattern such as periodic spikes, continuous drift, or repeated occurrence of specific amplitudes. At this time, the kernel density distribution is concentrated in a few deviation value intervals, the probability distribution is extremely uneven, and the calculated BHE value is significantly reduced.

[0118] when A value greater than 3 and less than or equal to 4.5 indicates that the blade has a micro-crack. A value greater than 4.5 indicates severe structural damage to the blade. A value of 3 or less indicates that the leaf blades are normal.

[0119] By setting threshold ranges of 3 and 4.5, this step discretizes continuous health entropy values ​​into three categories with clear engineering significance: normal, microcracks, and severe damage. Specifically, a BHE greater than 4.5 corresponds to a high-entropy state, characterized by strong signal randomness and no deterministic damage pattern; a BHE between 3 and 4.5 corresponds to a decrease in entropy value, characterized by the deviation coefficients beginning to show a clustering trend, corresponding to microcracks; and a BHE less than or equal to 3 corresponds to a low-entropy state, characterized by an extremely concentrated distribution of deviation coefficients and a highly deterministic damage pattern, corresponding to severe structural damage with a length greater than 2 mm.

[0120] Please see Figure 3 The present invention also provides a wind turbine blade performance diagnostic device based on field operation data. The diagnostic device is used to perform the above-described diagnostic method, including:

[0121] The data acquisition module is used to collect the strain time series data and sound pressure data of each blade of the wind turbine to be diagnosed in real time. It performs Fourier transform on the sound pressure data to obtain the linear spectrum and determines the power spectrum corresponding to the linear spectrum. Based on the power spectrum, it compresses to generate a grayscale image, simultaneously acquires the radial wind speed of the blade, calculates the angle of attack at different positions of the blade, and uses the weighted average method to obtain the overall angle of attack deviation of the blade.

[0122] The fusion vector acquisition module is used to extract features from strain time series data to obtain time series vectors and spatial vectors. It uses a gated attention weight dynamic allocation method to dynamically allocate the weights of time series vectors, spatial vectors and angle of attack deviations. Based on the set weight coefficients, the time series vectors, spatial vectors and angle of attack deviations are concatenated to obtain the fusion vector of the current state of the blade.

[0123] The deviation coefficient calculation module is used to compare the fusion vector of the healthy wind turbine blade sample with the healthy reference vector using Mahalanobis distance to obtain the deviation coefficient of the blade. Taking the current state time as the starting point, a time window and sampling frequency are set, and the deviation coefficient is sampled within the time window. The kernel density of each sampling point is calculated using a Gaussian kernel.

[0124] The comparison and classification module is used to calculate the health entropy value of the leaf within each time window based on the kernel density of each sampling point using the information entropy method, and to set the threshold range of the entropy value to classify the health entropy value of the leaf.

[0125] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0126] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0127] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0128] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for diagnosing the blade performance of wind turbines based on field operation data, characterized by the following steps: include: S1: Real-time acquisition of strain time series data and sound pressure data of each blade of the wind turbine to be diagnosed, Fourier transform of the sound pressure data to obtain a linear spectrum and determine the power spectrum corresponding to the linear spectrum, compress the power spectrum to generate a grayscale image, synchronously acquire the radial wind speed of the blade, calculate the angle of attack at different positions of the blade and use the weighted average method to obtain the overall angle of attack deviation of the blade. S2: Feature extraction is performed on the strain time series data to obtain time series vector and spatial vector. The gating attention weight dynamic allocation method is used to dynamically allocate the weights of the time series vector, spatial vector and angle of attack deviation. Based on the set weight coefficients, the time series vector, spatial vector and angle of attack deviation are concatenated to obtain the fusion vector of the current state of the blade. S4: Based on the fusion vector of healthy wind turbine blade samples as the health reference vector, the fusion vector of the current state of each blade and the health reference vector are compared using the Mahalanobis distance method to obtain the deviation coefficient of the blade. Taking the current state time as the starting point, a time window and sampling frequency are set, and the deviation coefficient is sampled within the time window. The kernel density of each sampling point is calculated using a Gaussian kernel. S5: Based on the kernel density of each sampling point, the health entropy value of the leaf within each time window is calculated using the information entropy method, and a threshold range for the entropy value is set to classify the health entropy value of the leaf.

2. The wind turbine blade performance diagnosis method based on field operation data according to claim 1, characterized in that, The specific steps to generate a grayscale image are as follows: Sound pressure data is collected from the transition region of the blade from the root to the airfoil section. The sound pressure data is sampled at a preset sampling rate and processed in frames. Fourier transform is performed on each frame to obtain the linear amplitude spectrum of each frame. The power spectrum is obtained by squaring the linear amplitude spectrum. The energy of each frequency point in the power spectrum is multiplied by the response value of a Mel filter of preset dimension. The frequency range of the Mel filter covers 0 to half of the preset sampling rate. The results of each filter are then summed to obtain the power spectrum of the preset dimension Mel band. The power spectrum of each Mel band is compressed by taking the natural logarithm and normalized by Z-score. The normalized power spectrum is linearly mapped to the grayscale range of 0 to 255. A time window containing a preset number of frames is set and updated at a preset time frequency. A column of grayscale values ​​of preset dimension is generated for each frame. The time axis is stitched together along the horizontal direction to finally form a grayscale image with the preset number of frames as rows and the preset dimension as columns.

3. The wind turbine blade performance diagnosis method based on field operation data according to claim 1, characterized in that, The weighted average method is used to calculate the overall angle of attack deviation of the blades. The specific steps are as follows: The radial wind speed of the blades is acquired in real time, and the angle of attack deviation is calculated using the following formula: Indicates the axial wind speed of the blades; Indicates the blade rotational speed; Indicates the radius of rotation of the blade; Indicates the pitch angle; The radius of rotation of the blade is indicated by Angle of attack deviation; The angle of attack at 25%, 50%, and 75% of the blade radius is calculated in real time, and a weighted average is calculated as the overall angle of attack deviation of the blade. in, This indicates the real-time overall angle of attack deviation of the blade.

4. The wind turbine blade performance diagnosis method based on field operation data according to claim 1, characterized in that, The specific steps for extracting time series vectors are as follows: The location of the main beam axis at which the historical blade is obtained is at a distance from the hub. Strain time-series data samples were collected at the location, and the health status of the blades was classified by category labeling. Undamaged blades were marked as normal, and blades with crack lengths ranging from... The blades were marked as microcracks, with crack lengths greater than [a certain value]. Leaves are marked as severely damaged; the category labels are converted to One-Hot encoding, specifically: normal leaves are marked as... Microcracks are Severe damage is ,in Indicates the total length of the leaf; Set a sliding time window and a number of time steps, moving at fixed time intervals. Add a classification head to the classifier, which will use the strain time series data samples within the sliding time window as... The input will The predicted vector output from the hidden layer is mapped to the corresponding label probability. The label probability is compared with the corresponding One-Hot encoding, the cross-entropy loss is calculated, and the loss is adjusted inversely. The parameters of the classification head are set, and training is completed when the loss entropy is less than a preset value; Training completed Remove the classification head in the middle, and position the main beam axis of the blade at a distance from the hub. Real-time strain time-series data of the location is input into the trained system. In the network, extract the time-series vector of the last time step within the window.

5. The wind turbine blade performance diagnosis method based on field operation data according to claim 2, characterized in that, To obtain the fusion vector of the blade in the current state, specifically: Input the grayscale image into the process of removing fully connected layers. In the network, the spatial vector is obtained. Based on the temporal vector and the spatial vector, the first, second, and third weight coefficients are calculated using a gated attention weight dynamic allocation method, as follows: The temporal and spatial vectors are linearly projected through different weight matrices, and the projection results are summed to obtain a joint feature vector. A hyperbolic tangent function is then applied to the joint feature vector for activation, mapping the numerical range to... In the interval, the three preset attention weight vectors are used to perform dot product operations with the activated joint feature vector, and the result is then processed. The values ​​of the obtained vector elements are used as the first, second, and third weighting coefficients, respectively; The vector concatenation method is used to concatenate the time-series vector, spatial vector, and angle-of-attack deviation. Specifically, the first, second, and third weight coefficients are multiplied by the time-series vector, spatial vector, and angle-of-attack deviation, respectively. The product is then expanded into an equal-dimensional vector by padding, i.e., all padding values ​​are set to 0. The padded vectors are then concatenated sequentially by connecting the first and last ends to obtain the fusion vector of the blade in the current state.

6. The wind turbine blade performance diagnosis method based on field operation data according to claim 5, characterized in that, The deviation coefficient of the blade in its current state is calculated using the Mahalanobis distance method. The specific steps are as follows: Obtain the fusion vector of historical healthy wind turbine blade samples, calculate the mean vector of all samples as the health baseline vector, and calculate the deviation coefficient of the current state blades using the Mahalanobis distance method based on the health baseline vector and the fusion vector. Subtract the current state's fusion vector from the health baseline vector dimension by dimension to obtain the deviation for each dimension. Based on the deviation, calculate the covariance matrix of the fusion vector of the historical healthy wind turbine blade samples. Invert the covariance matrix and use the inverted matrix as the weight matrix. Weight the deviation for each dimension and take the square root of the weighted result to obtain the deviation coefficient.

7. The wind turbine blade performance diagnosis method based on field operation data according to claim 6, characterized in that, The kernel density of each sampling point is calculated using a Gaussian kernel. The specific steps are as follows: Starting from the current state, a time window and sampling frequency are set. Within the time window, the deviation coefficients are sampled, and the kernel density of each sampling point is calculated using a Gaussian kernel. in, Indicates the time window, the first Kernel density of each sampling point; Indicates the first Deviation coefficient for each sampling point; Indicates the time window, the first Deviation coefficient for each sampling point; Indicates the number of sampling points; Indicates Gaussian bandwidth; Based on the kernel density of each sampling point, the health entropy value within the leaf time window is calculated using the information entropy method, and a threshold range for the entropy value is set to classify the health entropy value: in, This represents the health entropy value within the leaf's time window; when A value greater than 3 and less than or equal to 4.5 indicates that the blade has a micro-crack. A value greater than 4.5 indicates severe structural damage to the blade. A value of 3 or less indicates that the leaf blades are normal.

8. A wind turbine blade performance diagnostic device based on field operation data, characterized in that: The diagnostic device is used to perform the diagnostic method according to any one of claims 1-7, including: The data acquisition module is used to collect the strain time series data and sound pressure data of each blade of the wind turbine to be diagnosed in real time. It performs Fourier transform on the sound pressure data to obtain the linear spectrum and determines the power spectrum corresponding to the linear spectrum. Based on the power spectrum, it compresses to generate a grayscale image, simultaneously acquires the radial wind speed of the blade, calculates the angle of attack at different positions of the blade, and uses the weighted average method to obtain the overall angle of attack deviation of the blade. The fusion vector acquisition module is used to extract features from strain time series data to obtain time series vectors and spatial vectors. It uses a gated attention weight dynamic allocation method to dynamically allocate the weights of time series vectors, spatial vectors and angle of attack deviations. Based on the set weight coefficients, the time series vectors, spatial vectors and angle of attack deviations are concatenated to obtain the fusion vector of the current state of the blade. The deviation coefficient calculation module is used to compare the fusion vector of the healthy wind turbine blade sample with the healthy reference vector using Mahalanobis distance to obtain the deviation coefficient of the blade. Taking the current state time as the starting point, a time window and sampling frequency are set, and the deviation coefficient is sampled within the time window. The kernel density of each sampling point is calculated using a Gaussian kernel. The comparison and classification module is used to calculate the health entropy value of the leaf within each time window based on the kernel density of each sampling point using the information entropy method, and to set the threshold range of the entropy value to classify the health entropy value of the leaf.

Images