Wind turbine blade damage diagnosis method and system based on deep learning

Abstract

The application discloses a wind turbine blade damage diagnosis method and system based on deep learning, and relates to the technical field of wind turbine blade damage diagnosis. The method comprises the following steps: acquiring vibration array signals and acoustic emission array signals of a blade and performing pretreatment to obtain time-synchronized vibration signal segments and acoustic emission signal segments; inputting the two kinds of signal segments into a feature extractor constructed through a confrontation training mechanism to extract vibration damage features and acoustic emission damage features irrelevant to working conditions; performing micro-crack detection by using the acoustic emission damage features, obtaining a spatial clue vector from the acoustic emission damage features when detection is performed; and performing spatial attention weighting on the vibration damage features according to the spatial clue vector to obtain a spatial attention weight map, and further obtaining a blade damage position. The application effectively eliminates working condition interference through the confrontation training mechanism, combines the high sensitivity of acoustic emission signals and the accurate positioning capability of vibration signals, and realizes early detection and accurate positioning.

Description

Technical Field

[0001] This invention relates to the field of wind turbine blade damage diagnosis technology, specifically to a deep learning-based method and system for wind turbine blade damage diagnosis. Background Technology

[0002] Wind turbine blades are subjected to complex alternating loads and harsh environments over long periods of time, making them highly susceptible to damage such as cracks and peeling, which seriously affects the safety and power generation efficiency of the unit.

[0003] Existing deep learning-based blade damage diagnosis methods typically rely on input information such as images, vibration, and acoustic signatures. Image-based methods are unable to perform real-time monitoring due to the difficulty in acquiring clear images during operation. Vibration or acoustic signature information can be used for line monitoring. Vibration signals are sensitive to changes in structural stiffness and can reflect the overall degree of damage; acoustic emission signals are highly sensitive to the initiation and propagation of microcracks, but are easily affected by noise. Existing methods usually simply fuse the features of these two signals for diagnosis or combine the results of individual detections, failing to fully utilize the sensitivity differences between these two signals at different stages of damage evolution and their inherent physical connections. This results in poor physical interpretability of the diagnostic results and an inability to locate the spatial distribution of damage on the blade surface, making it difficult to provide decision support for maintenance.

[0004] In addition, the operating status of the wind turbine itself and environmental noise can significantly change the statistical characteristics of the collected signals, causing the performance of models trained on laboratory data to drop sharply in field applications, greatly reducing the effectiveness of practical applications. Summary of the Invention

[0005] This invention provides a method and system for diagnosing wind turbine blade damage based on deep learning, in order to solve the problem that existing wind turbine blade diagnosis methods are difficult to accurately locate the damage location in practical applications.

[0006] This invention is achieved through the following technical solution:

[0007] A first aspect of the present invention provides a deep learning-based method for diagnosing damage to wind turbine blades, comprising:

[0008] Acquire the vibration array signal and acoustic emission array signal of the blade;

[0009] The vibration array signal and acoustic emission array signal are preprocessed to obtain time-synchronized vibration signal segments and acoustic emission signal segments;

[0010] The vibration signal fragment and acoustic emission signal fragment are input into a pre-constructed feature extractor to obtain vibration damage features and acoustic emission damage features. The feature extractor is obtained through an adversarial training mechanism, so that the features learned by the feature extractor have the ability to maximize damage discrimination performance and minimize working condition discrimination performance, so as to extract working condition-independent damage features.

[0011] Microcrack detection is performed using the acoustic emission damage features. If a microcrack is detected, a spatial cue vector is obtained from the acoustic emission damage features.

[0012] The vibration damage features are spatially attention-weighted based on the spatial cue vector to obtain a spatial attention weight map, and the blade damage location is obtained based on the spatial attention weight map.

[0013] Furthermore, the feature extractor adopts a dual-branch shared underlying structure, including:

[0014] Vibration branch, used to extract vibration damage features from the vibration signal segment;

[0015] Acoustic emission branch, used to extract acoustic emission impairment features from the acoustic emission signal segment;

[0016] A shared fusion layer is used to stitch together the vibration damage features and acoustic emission damage features to obtain fused damage features;

[0017] During adversarial training, the fused damage features are used by the damage discriminator to perform damage discrimination tasks and the operating condition discriminator to perform operating condition discrimination tasks. The damage discriminator is used to determine the damage probability based on the fused damage features, and the operating condition discriminator is used to determine the operating condition information of the wind turbine based on the fused damage features.

[0018] Furthermore, the adversarial training method includes:

[0019] Build a damage discriminant and a working condition discriminant;

[0020] The damage discriminator takes the fused damage features as input and the reconstruction error of the lossless signal as the supervision target, aiming to minimize the signal reconstruction loss. Conduct training;

[0021] The operating condition discriminator takes the fused damage features as input and the operating condition information of the wind turbine as the monitoring target, and minimizes the operating condition discrimination loss. Conduct training;

[0022] Constructing total loss The feature extractor is trained by minimizing the total loss.

[0023] Furthermore, the operating condition information includes at least one of the following: fan speed, ambient wind speed, and ambient noise.

[0024] Furthermore, the step of performing microcrack detection using the acoustic emission damage characteristics includes:

[0025] The acoustic emission damage characteristics are input into the crack detection subnetwork to obtain the crack initiation probability.

[0026] When the probability of crack initiation is greater than a preset threshold, it is determined that a microcrack exists.

[0027] Further, obtaining the spatial cue vector from the acoustic emission damage features includes:

[0028] Extract the acoustic emission array signal within the time window corresponding to the acoustic emission damage characteristics to obtain the target acoustic emission array signal;

[0029] Based on the arrival time difference of each acoustic emission signal received by each acoustic emission sensor in the target acoustic emission array signal, the rough spatial coordinates of the damage source are calculated.

[0030] The rough spatial coordinates are encoded into spatial cue vectors of a predetermined dimension.

[0031] Further, the step of spatially attention-weighting the vibration damage features based on the spatial cue vector to obtain a spatial attention weight map includes:

[0032] The spatial clue vector and the vibration damage features are fused to obtain the fused features;

[0033] The fused features are input into an attention generation network; the attention generation network includes at least one fully connected layer for outputting an initial weight vector with the same spatial dimension as the vibration damage features.

[0034] A normalization function is applied to the initial weight vector to obtain a spatial attention weight map; wherein each element value in the spatial attention weight map represents the damage probability at the corresponding spatial location on the blade.

[0035] Furthermore, the method also includes:

[0036] The suspected damage area is determined based on the spatial attention weight map;

[0037] The target vibration array signal is obtained by calling the vibration array signal within the time window corresponding to the vibration damage characteristics;

[0038] A beamforming algorithm is applied to scan the target vibration array signal within the suspected damage area to obtain the location of blade damage.

[0039] Furthermore, the application of a beamforming algorithm to scan the target vibration array signal within the suspected damage area to obtain the blade damage location includes:

[0040] Obtain the theoretical signal propagation delay from each discrete position on the blade to each vibration sensor;

[0041] For each candidate location within the suspected damage area, the vibration signals received by each vibration sensor are phase-compensated and aligned according to the theoretical signal propagation delay, and the energy response value of the candidate location is calculated.

[0042] The candidate location with the largest energy response value is selected as the blade damage location.

[0043] A second aspect of the present invention provides a deep learning-based wind turbine blade damage diagnosis system, comprising:

[0044] The data acquisition module is used to acquire the vibration array signal and acoustic emission array signal of the blade;

[0045] The preprocessing module is used to preprocess the vibration array signal and the acoustic emission array signal to obtain time-synchronized vibration signal segments and acoustic emission signal segments;

[0046] The feature decoupling module includes a pre-built feature extractor for extracting features from the vibration signal segment and the acoustic emission signal segment to obtain vibration damage features and acoustic emission damage features that are independent of the working condition. The feature extractor is obtained through an adversarial training mechanism, so that the features learned by the feature extractor have the ability to maximize damage discrimination performance and minimize working condition discrimination performance.

[0047] The crack detection module is used to perform microcrack detection using the acoustic emission damage features. If a microcrack is detected, a spatial cue vector is obtained from the acoustic emission damage features.

[0048] The crack location module is used to perform spatial attention weighting on the vibration damage features based on the spatial cue vector to obtain a spatial attention weight map, and to obtain the blade damage location based on the spatial attention weight map.

[0049] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0050] 1. High sensitivity and accuracy: It utilizes the high sensitivity of acoustic emission signals to capture early micro-cracks, and then uses vibration signals to achieve precise positioning and quantitative assessment, thus realizing damage detection and damage localization;

[0051] 2. Strong physical interpretability: The design of the diagnostic model closely follows the physical mechanism of blade damage (acoustic emission response to crack initiation, vibration response to structural stiffness changes), the diagnostic process has clear physical logic, and the results are more reliable;

[0052] 3. High environmental robustness: Through the adversarial feature decoupling mechanism, the interference of changes in operating conditions such as wind speed and rotation speed on the diagnostic results is effectively eliminated, enabling the model to maintain stable performance in complex and ever-changing field environments. Attached Figure Description

[0053] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0054] Figure 1 This is a flowchart of a leaf damage diagnosis method according to an embodiment of the present invention;

[0055] Figure 2 This is a structural diagram of an adversarial training network according to an embodiment of the present invention;

[0056] Figure 3 This is a flowchart of a method for locating the damage site according to an embodiment of the present invention;

[0057] Figure 4 This is a block diagram of a leaf damage diagnosis system according to an embodiment of the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0059] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification, claims, and accompanying drawings of this invention are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to other steps or units inherent in the device.

[0060] The terminology used in the various embodiments of the invention is for the purpose of describing particular embodiments only and is not intended to limit the various embodiments of the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which the various embodiments of the invention pertain. The terms (such as those defined in commonly used dictionaries) are to be interpreted as having the same meaning as in the context of the relevant technical field and are not to be interpreted as having an idealized or overly formal meaning, unless clearly defined in the various embodiments of the invention.

[0061] Please see Figure 1 The present invention proposes a deep learning-based method for diagnosing wind turbine blade damage, comprising the following steps:

[0062] S101, acquire the vibration array signal and acoustic emission array signal of the blade.

[0063] In actual diagnostics, vibration array signals and acoustic emission array signals of the wind turbine blades during operation are acquired. The vibration array signals are obtained through a distributed array of vibration sensors deployed on the blades. Specifically, multiple vibration sensors can be arranged along the blade span at key locations such as the root, middle, and tip to collect vibration signals from different positions on the blade surface, forming a multi-channel vibration array signal.

[0064] Vibration sensors can be accelerometers or fiber optic grating sensors. For the same wind turbine blade, vibration sensors of the same model and parameter settings are deployed. Each sensor continuously collects vibration waveform data of the blade during operation synchronously based on a time synchronization system at the same sampling frequency, and sends the data to the diagnostic system at the same time nodes and time intervals.

[0065] The acoustic emission array signal is acquired by an array of acoustic emission sensors deployed on the blade. Acoustic emission signals are sensitive to the initiation of microcracks and can capture the high-frequency elastic waves released during microcracks and crack propagation. Since acoustic emission signals attenuate rapidly when propagating in composite materials, and the blade root is a high-risk area for damage, acoustic emission sensors can be placed at the root of each blade.

[0066] For each wind turbine to be diagnosed, a separate signal acquisition system consisting of M distributed vibration sensor arrays and N acoustic emission sensor arrays is deployed. The diagnostic system assigns a unique number to each acquisition system and wind turbine, collecting, storing, and differentiating these data using different numbers to achieve data isolation and individual diagnosis. The system separately stores the M channels of vibration array signals and the N channels of acoustic emission array signals for each wind turbine, and may also include 1- to n-dimensional operating condition information characterizing the wind turbine's operating status.

[0067] In one specific implementation, a two-dimensional or three-dimensional coordinate system is pre-established with the center of each wind turbine blade as the origin. The positions of the sensors arranged on this system are encoded into the coordinate system, with each sensor position corresponding one-to-one with the blade number and the discrete spatial positions of the blade, serving as the basis for subsequent damage localization. In practical applications, a two-dimensional coordinate system is preferred. Based on the wind turbine physical model, a mapping relationship between two-dimensional and three-dimensional coordinates can be established, mapping each sensor and its discrete position on the blade to a two-dimensional plane. Using two-dimensional coordinates for localization calculations simplifies the calculation process.

[0068] To ensure the spatiotemporal consistency of multi-source data, all field sensors are synchronously acquired through a unified time synchronization system to ensure that vibration signals and acoustic emission signals have globally consistent timestamps.

[0069] In one implementation, the method further includes a step of collecting operating condition information of the wind turbine. This operating condition information includes the turbine's own operating status (such as turbine speed and power) and environmental parameters (such as wind speed and ambient background noise). This information can be acquired in real time through the unit's SCADA system or on-site monitoring equipment, and time synchronization is ensured via GPS. The synchronously collected vibration array signals, acoustic emission array signals, and operating condition information of the wind turbine blades can be used to train a feature extraction network to extract damage features unrelated to the operating conditions from the vibration array signals and acoustic emission array signals.

[0070] S102 preprocesses the vibration array signal and acoustic emission array signal to obtain time-synchronized vibration signal segments and acoustic emission signal segments.

[0071] These preprocessing steps may include outlier removal, bandpass filtering, data slicing, and time alignment of the acquired signals. Outlier removal primarily eliminates abnormal data points caused by sensor malfunctions or strong external interference; by comparing data points, obvious anomalies can be identified. Bandpass filters are designed based on the frequency band characteristics of vibration and acoustic emission signals. Vibration signals typically focus on the low-frequency band, while acoustic emission signals focus on the high-frequency band.

[0072] After the above processing, the acquired data undergoes time alignment and slicing. Using a fixed time window T (e.g., 10 seconds), the continuously acquired signals are sliced, dividing the vibration array signal and acoustic emission array signal into data segments of the same time length. Data within each time window constitutes a data sample. Based on a unified timestamp, it is ensured that the vibration signal segment and acoustic emission signal segment contained in each sample strictly correspond in time.

[0073] Each vibration signal segment can be organized into a dimension of The matrix, where T is the number of sampling points within the time window. The number of vibration sensors. Each acoustic emission signal segment can be organized into a dimension of... The matrix, This represents the number of acoustic emission sensors.

[0074] S103, input the vibration signal segment and the acoustic emission signal segment into the pre-built feature extractor for feature extraction to obtain vibration damage features and acoustic emission damage features.

[0075] In this embodiment, the feature extractor employs a dual-branch feature extraction structure. For vibration features, a multi-layer large-kernel convolutional network is used, which can capture the low-frequency modal features of the vibration signal. For acoustic emission features, a multi-layer small-kernel convolutional network is used, as small convolutional kernels are more suitable for extracting the high-frequency transient features of the acoustic emission signal.

[0076] The feature extractor is trained through an adversarial training mechanism, which enables it to learn features that maximize damage discrimination performance and minimize working condition discrimination performance, thereby extracting damage features that are independent of working conditions.

[0077] In one specific implementation, the adversarial training network employs, for example... Figure 2 The training process for the structure shown includes:

[0078] Pre-build feature extraction network Damage discriminator and operating condition discriminator Among them, feature extractor A dual-branch feature extraction structure is adopted, including a vibration branch. Harmony emission branch Vibration branch From vibration signal segments Extracting vibration damage features Acoustic emission branch Extracting acoustic emission damage features from acoustic emission signal fragment Ga Damage discriminator The aim is to determine the probability of damage from the features learned by the feature extractor, while the condition discriminator... The aim is to identify operating condition information (such as one or more of the main influencing parameters being ambient wind speed, ambient noise, fan speed, or fan frequency) from the features learned by the feature extractor.

[0079] During adversarial training, the damage discriminator and operating condition discriminator The inputs are the same, representing vibration damage characteristics. Harmony and acoustic emission damage characteristics splicing and fusion features The damage discriminator D1 uses the reconstruction error of the lossless signal as its monitoring target, and aims to minimize the signal reconstruction loss. Training; Operating condition discriminator Using operating condition information as the monitoring target, losses are identified by minimizing operating conditions. Conduct training.

[0080] Feature extractor The training objective is to improve the damage discriminator. Maximize the discrimination performance of the operating condition discriminator The discrimination performance is minimized. Therefore, the total loss function can be set as:

[0081]

[0082] in, For the signal reconstruction loss of the damage discriminator, For the loss of the operating condition discrimination device, This is a balancing coefficient used to adjust the relative weights of damage discrimination loss and operating condition discrimination loss in the total loss. The value range is 0.1 to 1.0. When the value is small, the feature extractor focuses more on damage detection performance; when... When the value is large, more emphasis is placed on the decoupling effect of operating conditions. In practical applications, The value can be 1, or the optimal value can be determined through cross-validation. value.

[0083] To minimize total loss The parameters of the feature extractor are trained to obtain pure, condition-independent vibration damage features and acoustic emission damage features.

[0084] In one specific implementation, the feature extractor adopts a dual-branch shared underlying structure, including a vibration branch, an acoustic emission branch, and a shared fusion layer. The vibration branch and the acoustic emission branch share a common fusion layer, which splices and fuses the vibration damage features and acoustic emission damage features to obtain fused damage features, which are then input into the damage discriminator and the operating condition discriminator, respectively.

[0085] In one specific implementation, the method of training a feature extraction network using an adversarial training mechanism includes:

[0086] S301, Sample Data Acquisition and Preparation.

[0087] The training data includes vibration array signals, acoustic emission array signals, and operating condition information from multiple wind turbine samples. This acquired information is preprocessed to obtain time-synchronized vibration signal segments, acoustic emission signal segments, and operating condition feature vectors. The preprocessing also includes outlier removal, bandpass filtering, data slicing, and time alignment of the time-series data of the vibration array signals, acoustic emission array signals, and operating condition information.

[0088] The operating condition feature vector is obtained by encoding the operating condition information and is a 1 to n-dimensional feature vector, with each dimension corresponding to an operating condition parameter. This operating condition information includes at least one of the following: fan speed, ambient wind speed, and ambient noise.

[0089] S302 employs an alternating training strategy to train the feature extractor, damage discriminator, and operating condition discriminator:

[0090] S3021: First, fix the model parameters of the feature extractor, and train the damage discriminator and the working condition discriminator to achieve their optimal performance on their respective tasks.

[0091] S3022, fix the model parameters of the damage discriminator and the working condition discriminator, and train the feature extractor so that the extracted features can maximize the performance of the damage discriminator and minimize the working condition discriminator (even if the total loss is low). Minimum);

[0092] S303, execute S3021~S3022 alternately multiple times until each model converges.

[0093] In one implementation, the damage discriminator employs an autoencoder structure, whose input is fused damage features and whose output is a reconstructed lossless signal template. The autoencoder consists of an encoder and a decoder, and is trained by minimizing the reconstruction error between the input and output. For healthy samples, the reconstruction error is small; for damaged samples, the reconstruction error increases significantly. The magnitude of the reconstruction error can be directly used as a measure of the degree of damage.

[0094] In one implementation, the operating condition discriminator employs a multilayer perceptron structure, with fused damage features as input and predicted operating condition information (such as wind speed, rotational speed, etc.) as output. The operating condition discriminator is trained by minimizing the mean square error between the predicted value and the actual operating condition.

[0095] Through the above adversarial training, the feature extractor can extract condition-independent damage features from the original vibration and acoustic emission features, and fully combines the characteristics of both types of features to improve the reliability of damage diagnosis. During online diagnosis, only the vibration and acoustic emission signal segments acquired in real time need to be input into the feature extractor to filter out interfering features.

[0096] S104 utilizes acoustic emission damage features to perform microcrack detection. If a microcrack is detected, a spatial cue vector is obtained from the acoustic emission damage features.

[0097] In this implementation, microcrack detection is first performed using acoustic emission damage features, followed by damage localization. Because acoustic emission signals are highly sensitive to microcrack initiation and propagation, they can capture elastic waves released from minute damage within the structure. Therefore, acoustic emission damage features contain rich early damage information. By identifying the moment of crack initiation from the acoustic emission damage features through a crack detection sub-network, timely warnings can be issued when the damage is still in its early stages, providing a trigger signal for subsequent precise localization. This step fully leverages the high sensitivity of acoustic emission signals, complementing the subsequent vibration signal-based localization step: acoustic emission signals handle sensing, while vibration signals handle localization.

[0098] Specifically, acoustic emission damage features are input into a crack detection sub-network, which is a binary classification neural network, employing a multilayer perceptron or temporal convolutional network structure. Its output is a scalar value representing the probability of microcrack initiation activity at the current moment. When the crack initiation probability exceeds a preset threshold (e.g., 0.8), a microcrack is detected, immediately triggering the subsequent damage localization process.

[0099] The crack detection subnetwork employs supervised training, with training samples consisting of historically acquired acoustic emission damage features, labeled with the presence or absence of a crack at the corresponding moment. After training, the crack detection subnetwork can output the probability of a crack existing at the current moment in real time. The samples and their labels can be obtained through offline detection or simulation data collection.

[0100] When the probability of crack initiation exceeds a preset threshold, the generation of a spatial clue vector is triggered, which is obtained based on the acoustic emission damage characteristics. This spatial clue vector provides approximate information about the possible location of the damage and can be used to narrow down the search range for subsequent precise localization.

[0101] In one specific implementation, the method for obtaining spatial cue vectors based on acoustic emission damage characteristics includes:

[0102] S401, extract the acoustic emission array signal within the time window corresponding to the acoustic emission damage characteristics to obtain the target acoustic emission array signal;

[0103] S402, calculate the approximate spatial coordinates of the damage source based on the arrival time difference of each acoustic emission signal received by each acoustic emission sensor in the target acoustic emission array signal;

[0104] S403 encodes coarse spatial coordinates into spatial cue vectors of a predetermined dimension.

[0105] The acoustic emission array signal within the time window corresponding to the acoustic emission damage features refers to the original acoustic emission array signal within the same time window T used during feature extraction. Since the acoustic emission damage features are extracted from the acoustic emission signal segments within this time window, extracting the original signal from the same time window ensures the temporal consistency of the data.

[0106] The arrival time difference of the acoustic emission signals received by each acoustic emission sensor can be calculated by performing cross-correlation analysis on the signals of each channel. Specifically, the signal of a reference sensor (such as the first sensor) is selected as a benchmark, and the cross-correlation function between the signals of other sensors and the benchmark signal is calculated. The peak position of the cross-correlation function is the relative arrival time difference.

[0107] The approximate spatial coordinates of the damage source can be calculated based on the time difference of arrival. Classical acoustic emission localization methods, such as the time difference localization method based on the hyperbolic principle, can be used. Given the spatial coordinates of each sensor and the propagation speed of sound waves in the blade material, the approximate spatial coordinates of the damage source can be obtained by solving the equations for the time differences of arrival between multiple sensors.

[0108] The coarse spatial coordinates are encoded into spatial cue vectors of a predetermined dimension. Specifically, the blade is pre-divided into several grids in a two-dimensional coordinate system, and the grid positions where the coarse coordinates are located are encoded as one-hot vectors; or the coordinate values ​​are directly normalized and used as vector elements. In this embodiment, after the coarse spatial coordinates (x, y) are directly normalized to the [0, 1] interval, a series of spatial cue vectors c = [x_norm, y_norm] can be obtained, which contain the approximate location information of the damage.

[0109] S105, spatial attention weighting is applied to the vibration damage features based on the spatial cue vector to obtain a spatial attention weight map, and the blade damage location is obtained based on the spatial attention weight map.

[0110] In this step, the vibration damage features are derived from the vibration sensor array, reflecting information on changes in the structural stiffness of various parts of the blade. However, vibration signals are sensitive to global structural changes, making it difficult to directly locate minor local damage. By weighting the vibration damage features using the approximate damage location information provided by spatial cue vectors, the subsequent localization process can focus on the area indicated by the cue vectors, thereby improving localization accuracy and efficiency. This step combines the coarse location information of acoustic emission with the high-precision vibration localization capability, achieving collaborative diagnosis guided by acoustic emission signals and focused on vibration signals.

[0111] In one specific implementation, the spatial attention weight map is calculated as follows:

[0112] S501, the spatial clue vector and vibration damage features are fused to obtain the fused features;

[0113] S502, the fused features are input into the attention generation network; wherein, the attention generation network includes at least one fully connected layer for outputting an initial weight vector with the same spatial dimension as the vibration damage features;

[0114] S503, apply a normalization function to the initial weight vector to obtain the spatial attention weight map; where each element value in the spatial attention weight map represents the damage probability at the corresponding spatial location on the blade.

[0115] The spatial attention weight map A is a weight vector with the same spatial dimension as the vibration damage feature, where each element corresponds to a vibration sensor location. Since there is a fixed mapping relationship between sensor locations and blade spatial locations (determined through a pre-established blade coordinate system), each element in A can be directly mapped to a corresponding spatial location on the blade, representing the probability of damage at that location. The spatial attention weight map A reflects the distribution of damage probabilities in the blade space; regions with higher probability values ​​are more likely to have damage.

[0116] Through the above steps, the technical solution of this invention achieves the following technical effects: First, by using an adversarial training mechanism, pure damage features independent of the operating conditions are extracted from the original signal, solving the problem of environmental interference; second, high-sensitivity detection of microcracks is achieved using acoustic emission damage features; finally, by using spatial attention weighting, the coarse location information of acoustic emission is fused into the vibration features, realizing the probability distribution estimation of the damage location. The entire solution fully leverages the respective advantages of acoustic emission and vibration signals, achieving early detection and location estimation of microcracks under complex operating conditions.

[0117] Specifically, the structure of the attention generation network is as follows:

[0118] Feature fusion: Combining vibration damage features Fv with spatial cue vectors The fusion is performed to obtain the fusion feature z1;

[0119] Weight calculation: Input the fused feature z1 into one or more fully connected layers, and output a weight vector with the same dimensions as the Fv space;

[0120] Spatial weighting: Normalizing the weight vector using the Sigmoid function yields values ​​within a certain range. Spatial attention weight map between them.

[0121] Multiple fully connected layers in the attention generation network are sequentially connected to form a deep network structure. To enhance the network's expressive power, ReLU activation functions and Dropout layers can be added between the fully connected layers to prevent overfitting. The last layer outputs an initial weight vector with the same dimension as the vibration damage feature space.

[0122] In one implementation, the method further includes a step of determining the precise location of blade damage based on a spatial attention weight map, such as... Figure 3 As shown.

[0123] S601, firstly based on the spatial attention weight map Identify the suspected damage area R;

[0124] S602, call the vibration array signal within the time window corresponding to the vibration damage feature Fv to obtain the target vibration array signal;

[0125] S603 uses a beamforming algorithm to scan the target vibration array signal within the suspected damage area to obtain the location of blade damage.

[0126] Among them, the area of ​​suspected damage is defined. Spatial attention weight map The value of the middle element is greater than The area, namely ,in The attention threshold is defined. The vibration damage feature Fv is spatially weighted using the spatial cue vector c, thus enhancing the damaged area and weakening the non-damaged area. When The value of the middle element is greater than When this occurs, it indicates that the location has a high probability of damage. The setting can be adjusted according to actual monitoring needs, and is usually taken as a value between 0.7 and 0.9. In this embodiment, it is set to 0.8, that is, in the spatial attention weight map A, the area with a weight value greater than 0.8 is defined as the suspect area R.

[0127] By identifying the suspected damage area, the scanning range of the localization algorithm is limited to within the suspected area R. Specifically, the beamforming algorithm needs to calculate the energy response value for each candidate location within the suspected area. By limiting the candidate locations to R instead of the entire blade, the computational load is significantly reduced.

[0128] The beamforming algorithm used in this embodiment is a spatial filtering technique based on array signal processing. Its basic principle is to compensate for the delay of signals from different sensors and then superimpose them to coherently enhance signals from a specific direction while weakening signals from other directions. In this invention, beamforming is used to locate the spatial coordinates of the sound source (i.e., the damage location) on the blade.

[0129] In one specific implementation, the method for locating damage based on beamforming algorithms is as follows:

[0130] S701, obtain the theoretical signal propagation delay from each discrete position on the blade to each vibration sensor;

[0131] S702, for each candidate location within the suspected damage area, the vibration signals received by each vibration sensor are phase-compensated and aligned according to the theoretical signal propagation delay, and the energy response value of the candidate location is calculated.

[0132] S703 selects the candidate location with the largest energy response value as the blade damage location.

[0133] A pre-established database of propagation time delays for the blade structure is used, which records the theoretical signal propagation time delays from each discrete position on the blade to each vibration sensor. Regarding the suspected area Each candidate position within Based on the theoretical signal propagation time delay, the original vibration signal Perform phase compensation and alignment, and calculate the energy response value at that location:

[0134]

[0135] in, Indicates time, Indicates the sensor number, This represents the number of vibration sensors.

[0136] Traversing the entire suspected area energy response value The largest candidate location was determined as the precise spatial coordinates of the damage source, as shown below:

[0137]

[0138] Through the above steps, the final output is This provides the precise spatial coordinates of the damage source on the blade, enabling accurate location of the damage.

[0139] In a preferred embodiment, the method for scanning the target vibration array signal within the suspected damage area to obtain the blade damage location includes:

[0140] S801, spatial attention weight map Element-wise multiplication with the vibration damage feature Fv yields the focused vibration feature;

[0141] S802, call the vibration array signal within the time window corresponding to the vibration damage characteristics to obtain the target vibration array signal;

[0142] S803 performs spatial filtering on the target vibration array signal based on the focused vibration characteristics, and applies a beamforming algorithm to scan the original filtered vibration array signal within the suspected area to locate the damage location.

[0143] This embodiment constitutes a better localization implementation method. Step S801 multiplies the spatial attention weight map A with the vibration damage feature Fv element by element to obtain the focused vibration feature. This operation achieves spatial attention at the feature level. In the process, vibration characteristics in high-probability regions are preserved, while vibration characteristics in low-probability regions are suppressed.

[0144] Since the original vibration signal is the superposition of vibration responses at all locations on the blade, directly using beamforming would simultaneously receive signals from both inside and outside the suspected area, potentially introducing interference. By constructing a spatial filter using the focused vibration characteristics, signal components from the suspected area can be enhanced and signal components from outside the suspected area can be suppressed at the signal level, thereby improving the signal-to-noise ratio and positioning accuracy of subsequent beamforming.

[0145] Specifically, the vibration characteristics after focusing The system provides weight information for each spatial location, which can be used to construct a spatial filter. For each vibration sensor channel, the filter is adjusted based on the coupling relationship between the sensor and each spatial location. Given the weights, calculate the filter coefficients for that channel on the original signal. Weighting is performed to obtain the enhanced signal. , Indicates time, Indicates the sensor number.

[0146] Represented as:

[0147]

[0148] This operation is equivalent to spatially weighting the vibration characteristics, which preserves and amplifies signal components from high-probability regions, while suppressing signal components from low-probability regions.

[0149] In one specific implementation, a delay-and-sum beamforming method is used in the suspected area. The beamforming algorithm is used for scanning and positioning. Details are as follows:

[0150] S601, obtain the theoretical signal propagation delay from each discrete position on the blade to each vibration sensor. ;

[0151] S602, regarding the suspected area Each candidate position within Based on the time delay database, the enhanced signal Perform phase compensation and alignment, and calculate the energy response value at that location. ;

[0152] S603, traversing the entire suspected area energy response value The largest candidate location was determined as the precise spatial coordinates of the damage source.

[0153] In a preferred embodiment, the present invention may further include the step of modeling and predicting the spatiotemporal evolution of damage, which can predict the crack development trend based on historical diagnostic results.

[0154] First, a spatiotemporal graph convolutional network, such as the ST-GCN model, is constructed to model the evolution of damage. Using a graph sequence at each time step as input, and through alternating graph convolution and temporal convolution operations, the spatial propagation pattern and temporal accumulation trend of damage are captured, and the damage prediction for future time steps is output.

[0155] The blade structure is pre-discretized into a finite number of nodes to construct a blade diagram structure. Historical vibration damage characteristics of each node's spatial location are then used as the basis for this analysis. The sequence serves as the feature vector of the node; the edge weights between nodes can be dynamically defined based on the structural mechanical properties of the blade (such as distance and stiffness coupling) and the damage correlation between nodes in the historical positioning results.

[0156] The following methods can be used to train spatiotemporal graph convolutional networks:

[0157] (1) Training sample acquisition: The historical monitoring data is processed through steps S101 to S105 to obtain the damage location and damage probability distribution of each time window, which are used as training samples for ST-GCN, and the actual location, size and type of damage are labeled.

[0158] (2) Graph structure construction: Discretize the leaf into N nodes, and the feature vector of node i at time t is It consists of a sequence of historical vibration damage characteristics at the node location. The edge weights between nodes... The determination is based on the structural mechanical properties of the blade (such as the distance between nodes and the stiffness coupling coefficient) and the correlation of historical damage.

[0159] (3) Model Training: The spatiotemporal graph convolutional network consists of multiple spatiotemporal graph convolutional layers, each of which includes spatial graph convolution and temporal convolution. Spatial graph convolution is used to aggregate information from neighboring nodes, while temporal convolution is used to capture temporal dependencies. During training, the graph sequence G(t-K+1), ..., G(t) from the past K time steps is used as input, and the damage state (location, size) from the future L time steps is used as output. Mean squared error is used as the loss function, and the model parameters are optimized through backpropagation.

[0160] After training, this spatiotemporal graph convolutional network can be used for real-time damage evolution prediction. In application, damage features from the current and K historical time steps are input into the model, and the model outputs predictions of the damage state for the next L time steps, including:

[0161] Damage location prediction: Predicting the possible location of the damage source at a future time;

[0162] Damage size prediction: Predicting the growth trend of crack length and width;

[0163] Damage severity score: Outputs a continuous score in the range [0,1] to quantify the health status of the leaf.

[0164] Based on the forecast results, maintenance personnel can arrange maintenance plans in advance to achieve predictive maintenance.

[0165] Embodiments of the present invention also provide a deep learning-based wind turbine blade damage diagnosis system, such as... Figure 4 As shown, the system includes:

[0166] The data acquisition module 401 is used to acquire the vibration array signal and acoustic emission array signal of the blade;

[0167] The preprocessing module 402 is used to preprocess the vibration array signal and the acoustic emission array signal to obtain time-synchronized vibration signal segments and acoustic emission signal segments;

[0168] The feature decoupling module 403 includes a pre-built feature extractor for extracting features from vibration signal segments and acoustic emission signal segments to obtain vibration damage features and acoustic emission damage features that are independent of the working conditions. The feature extractor is obtained through an adversarial training mechanism, which enables the features learned by the feature extractor to maximize damage discrimination performance and minimize working condition discrimination performance.

[0169] The crack detection module 404 is used to perform microcrack detection using acoustic emission damage features. If a microcrack is detected, a spatial cue vector is obtained from the acoustic emission damage features.

[0170] The crack location module 405 is used to perform spatial attention weighting on the vibration damage features according to the spatial cue vector to obtain a spatial attention weight map, and to obtain the blade damage location according to the spatial attention weight map.

[0171] Furthermore, the feature extractor adopts a dual-branch shared underlying structure, including:

[0172] Vibration branch, used to extract vibration damage features from vibration signal segments;

[0173] The acoustic emission branch is used to extract acoustic emission damage features from acoustic emission signal segments;

[0174] A shared fusion layer is used to stitch together vibration damage features and acoustic emission damage features to obtain fused damage features.

[0175] During adversarial training, the fused damage features are used by the damage discriminator to perform damage discrimination tasks and the operating condition discriminator to perform operating condition discrimination tasks. The damage discriminator is used to determine the damage probability based on the fused damage features, and the operating condition discriminator is used to determine the operating condition information of the wind turbine based on the fused damage features.

[0176] Furthermore, the system also includes an adversarial training module for performing the adversarial training process. It employs an alternating training strategy: the damage discriminator takes the fused damage features as input and the reconstruction error of the lossless signal as the supervision target, minimizing the signal reconstruction loss. Training is performed; the operating condition discriminator takes fused damage features as input and wind turbine operating condition information as the supervision target, aiming to minimize the operating condition discrimination loss. Conduct training; establish total loss The feature extractor is trained by minimizing the total loss.

[0177] Furthermore, the crack detection module 404 includes a crack detection sub-network. By inputting acoustic emission damage features into the crack detection sub-network, the crack initiation probability is obtained. When the crack initiation probability is greater than a preset threshold, it is determined that a microcrack exists, and the crack localization module is triggered to perform the localization process.

[0178] Furthermore, the crack detection module 404 also includes a damage cue generation module, which is configured to generate a spatial cue vector based on acoustic emission damage characteristics when the crack initiation probability is greater than a preset threshold. Specifically, this includes:

[0179] The extraction submodule is used to extract the acoustic emission array signal within the time window corresponding to the acoustic emission damage characteristics, and obtain the target acoustic emission array signal.

[0180] The coarse estimation submodule is used to calculate the rough spatial coordinates of the damage source based on the arrival time difference of each acoustic emission signal received by each acoustic emission sensor in the target acoustic emission array signal.

[0181] The encoding module is used to encode coarse spatial coordinates into spatial cue vectors of a predetermined dimension.

[0182] Furthermore, the crack localization module 405 includes a spatial attention weighting submodule, which is configured to perform spatial attention weighting on the vibration damage features according to the spatial cue vector when the crack initiation probability is greater than a preset threshold, to obtain a spatial attention weight map. It also includes a damage localization submodule, which is configured to locate the blade damage location based on the spatial attention weight map.

[0183] Furthermore, the crack localization module 405 includes a coarse localization submodule and a fine localization submodule. The coarse localization submodule is configured to determine the suspected damage area based on a spatial attention weight map; the fine localization submodule includes:

[0184] The extraction submodule is used to retrieve the vibration array signal within the time window corresponding to the vibration damage characteristics to obtain the target vibration array signal;

[0185] The scanning submodule is used to scan the target vibration array signal within the suspected damage area to obtain the location of blade damage.

[0186] Embodiments of the present invention also provide an electronic device including a processor and a memory, wherein the number of processors may be one or more. The memory, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and modules. The processor executes various functional applications and data processing of the electronic device by running the software programs, instructions, and modules stored in the memory, thereby implementing the deep learning-based wind turbine blade damage diagnosis method of any of the above embodiments of the present invention.

[0187] The memory may primarily comprise a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a given function; the data storage area may store data created based on terminal usage. Furthermore, the memory may include high-speed random access memory (RAM) and non-volatile memory, such as at least one disk storage device, flash memory, or other non-volatile solid-state storage device. In some instances, the memory may further include memory remotely located relative to the processor, which can be connected to the electronic device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks (LANs), mobile communication networks, and combinations thereof.

[0188] Embodiments of the present invention also provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the deep learning-based wind turbine blade damage diagnosis method of any embodiment of the present invention.

[0189] The computer storage medium of this invention can be any combination of one or more computer-readable media. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0190] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of sending, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0191] Embodiments of the present invention also provide a computer program product that, when run on a computer, causes the computer to execute the deep learning-based wind turbine blade damage diagnosis method of any of the above embodiments of the present invention.

[0192] The above embodiments are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the appended claims.

Claims

1. A deep learning-based method for diagnosing wind turbine blade damage, characterized in that, include: Acquire the vibration array signal and acoustic emission array signal of the blade; The vibration array signal and acoustic emission array signal are preprocessed to obtain time-synchronized vibration signal segments and acoustic emission signal segments; The vibration signal fragment and acoustic emission signal fragment are input into a pre-constructed feature extractor to obtain vibration damage features and acoustic emission damage features. The feature extractor is obtained through an adversarial training mechanism, so that the features learned by the feature extractor have the ability to maximize damage discrimination performance and minimize working condition discrimination performance, so as to extract working condition-independent damage features. Microcrack detection is performed using the acoustic emission damage features. If a microcrack is detected, a spatial cue vector is obtained from the acoustic emission damage features. The vibration damage features are spatially attention-weighted based on the spatial cue vector to obtain a spatial attention weight map, and the blade damage location is obtained based on the spatial attention weight map.

2. The deep learning-based wind turbine blade damage diagnosis method according to claim 1, characterized in that, The feature extractor adopts a dual-branch shared underlying structure, including: Vibration branch, used to extract vibration damage features from the vibration signal segment; Acoustic emission branch, used to extract acoustic emission impairment features from the acoustic emission signal segment; A shared fusion layer is used to stitch together the vibration damage features and acoustic emission damage features to obtain fused damage features; During adversarial training, the fused damage features are used by the damage discriminator to perform damage discrimination tasks and the operating condition discriminator to perform operating condition discrimination tasks. The damage discriminator is used to determine the damage probability based on the fused damage features, and the operating condition discriminator is used to determine the operating condition information of the wind turbine based on the fused damage features.

3. The deep learning-based wind turbine blade damage diagnosis method according to claim 2, wherein the adversarial training method includes: Build a damage discriminant and a working condition discriminant; The damage discriminator takes the fused damage features as input and the reconstruction error of the lossless signal as the supervision target, aiming to minimize the signal reconstruction loss. Conduct training; The operating condition discriminator takes the fused damage features as input and the operating condition information of the wind turbine as the monitoring target, and minimizes the operating condition discrimination loss. Conduct training; Constructing total loss The feature extractor is trained by minimizing the total loss.

4. The deep learning-based wind turbine blade damage diagnosis method according to claim 3, characterized in that, The operating condition information includes at least one of the following: fan speed, ambient wind speed, and ambient noise.

5. The deep learning-based wind turbine blade damage diagnosis method according to claim 1, characterized in that, The method of performing microcrack detection using the acoustic emission damage characteristics includes: The acoustic emission damage characteristics are input into the crack detection subnetwork to obtain the crack initiation probability. When the probability of crack initiation is greater than a preset threshold, it is determined that a microcrack exists.

6. The deep learning-based wind turbine blade damage diagnosis method according to claim 1, characterized in that, The spatial cue vector obtained from the acoustic emission damage features includes: Extract the acoustic emission array signal within the time window corresponding to the acoustic emission damage characteristics to obtain the target acoustic emission array signal; Based on the arrival time difference of each acoustic emission signal received by each acoustic emission sensor in the target acoustic emission array signal, the rough spatial coordinates of the damage source are calculated. The rough spatial coordinates are encoded into spatial cue vectors of a predetermined dimension.

7. The deep learning-based wind turbine blade damage diagnosis method according to claim 6, characterized in that, The step of performing spatial attention weighting on the vibration damage features based on the spatial cue vector to obtain a spatial attention weight map includes: The spatial clue vector and the vibration damage features are fused to obtain the fused features; The fused features are input into an attention generation network; the attention generation network includes at least one fully connected layer for outputting an initial weight vector with the same spatial dimension as the vibration damage features. A normalization function is applied to the initial weight vector to obtain a spatial attention weight map; wherein each element value in the spatial attention weight map represents the damage probability at the corresponding spatial location on the blade.

8. The deep learning-based wind turbine blade damage diagnosis method according to claim 1, characterized in that, The method further includes: The suspected damage area is determined based on the spatial attention weight map; The target vibration array signal is obtained by calling the vibration array signal within the time window corresponding to the vibration damage characteristics; A beamforming algorithm is applied to scan the target vibration array signal within the suspected damage area to obtain the location of blade damage.

9. The deep learning-based wind turbine blade damage diagnosis method according to claim 8, characterized in that, The application of beamforming algorithm to scan the target vibration array signal within the suspected damage area to obtain the blade damage location includes: Obtain the theoretical signal propagation delay from each discrete position on the blade to each vibration sensor; For each candidate location within the suspected damage area, the vibration signals received by each vibration sensor are phase-compensated and aligned according to the theoretical signal propagation delay, and the energy response value of the candidate location is calculated. The candidate location with the largest energy response value is selected as the blade damage location.

10. A deep learning-based wind turbine blade damage diagnosis system, characterized in that, include: The data acquisition module is used to acquire the vibration array signal and acoustic emission array signal of the blade; The preprocessing module is used to preprocess the vibration array signal and the acoustic emission array signal to obtain time-synchronized vibration signal segments and acoustic emission signal segments; The feature decoupling module includes a pre-built feature extractor for extracting features from the vibration signal segment and the acoustic emission signal segment to obtain vibration damage features and acoustic emission damage features that are independent of the working condition. The feature extractor is obtained through an adversarial training mechanism, so that the features learned by the feature extractor have the ability to maximize damage discrimination performance and minimize working condition discrimination performance. The crack detection module is used to perform microcrack detection using the acoustic emission damage features. If a microcrack is detected, a spatial cue vector is obtained from the acoustic emission damage features. The crack location module is used to perform spatial attention weighting on the vibration damage features based on the spatial cue vector to obtain a spatial attention weight map, and to obtain the blade damage location based on the spatial attention weight map.

Images