A Wind Turbine Blade Condition Monitoring and Fault Early Warning System Based on Wavelet Clustering Algorithm

Abstract

This invention discloses a wind turbine blade condition monitoring and fault early warning system based on wavelet clustering algorithm, relating to the field of wind turbine blade fault diagnosis. The system synchronously collects blade operating data through vibration sensors and acoustic emission sensors, employs a multi-domain joint feature extraction method to fuse time-domain, frequency-domain, and time-frequency-domain features, and utilizes an improved wavelet clustering algorithm that automatically acquires local clustering information to perform unsupervised clustering analysis on the feature vectors, automatically identifying damage patterns. Fault classification is achieved by establishing a support vector machine diagnostic model, and an innovative acoustic emission event rate fluctuation verification mechanism based on the blade rotation periodicity is introduced, using vibration and acoustic emission signals to mutually verify the authenticity of the damage. This invention solves the problems of insufficient feature extraction, poor clustering effect, and high false alarm rate in traditional methods, achieving accurate identification and early warning of multiple damage patterns, effectively improving the accuracy of fault identification.

Description

Technical Field

[0001] This invention relates to the field of wind turbine blade fault diagnosis, specifically to a wind turbine blade condition monitoring and fault early warning system based on wavelet clustering algorithm. Background Technology

[0002] As a core component of wind turbine generators, the structural health of wind turbine blades directly affects the overall power generation efficiency, operational safety, and service life of the unit. Due to the long-term exposure to complex dynamic effects such as aerodynamic loads, gravitational loads, and environmental corrosion, blades are prone to microscopic damage such as matrix cracking, fiber breakage, and delamination. If these damages are not detected and addressed in a timely manner, they will gradually expand and lead to macroscopic structural failure, even resulting in catastrophic accidents such as blade breakage.

[0003] Currently, wind turbine blade condition monitoring mainly relies on methods such as vibration monitoring and acoustic emission monitoring. Vibration monitoring captures changes in the blade's dynamic response using accelerometers, while acoustic emission monitoring identifies early damage by capturing stress wave signals released from material damage.

[0004] However, these methods have the following problems in practical applications:

[0005] High signal processing difficulty: The blade operating environment is complex, and the collected signals are mixed with a large amount of noise, such as wind noise, mechanical vibration, and electromagnetic interference, making it difficult for traditional methods to effectively extract damage features. Insufficient feature representation: Existing methods mostly rely on single-domain features such as time or frequency domains, which cannot comprehensively represent the characteristics of the signal, resulting in low damage identification accuracy. Poor adaptability of clustering algorithms: Traditional clustering algorithms (such as K-means, DBSCAN, etc.) often fail to achieve ideal clustering results when processing high-dimensional, nonlinear, and non-Gaussian distributed wind turbine blade monitoring data. Difficulty in identifying multiple damage patterns: Blade damage takes many forms, including matrix cracking, fiber breakage, delamination, and other patterns, making it difficult for traditional methods to accurately identify and classify multiple damage patterns. Summary of the Invention

[0006] The purpose of this invention is to achieve accurate identification of wind turbine blade fault types.

[0007] To achieve the above objectives, this invention provides a wind turbine blade condition monitoring and fault early warning system based on wavelet clustering algorithm, the system comprising:

[0008] The calibration module is used to calibrate wind turbine blade specimens that are normal and undamaged as well as those with a known type of typical damage, and to obtain calibration results. The known typical damage includes type M.

[0009] The data acquisition module is used to acquire the original vibration signals and original acoustic emission signals of the wind turbine blade specimen under simulated operating conditions;

[0010] The feature extraction module is used to extract the time-domain and frequency-domain features of the original vibration signal, and to extract the time-frequency domain features of the original acoustic emission signal based on the first wavelet transform, and to fuse the time-domain features, the frequency-domain features and the time-frequency domain features to generate a fused feature vector;

[0011] The wavelet clustering module is used to perform cluster analysis on the fused feature vector based on the wavelet clustering algorithm to obtain N cluster categories, and assign a physical meaning to any cluster category based on the calibration result;

[0012] The fault diagnosis and early warning module is used for:

[0013] The time-domain features, the frequency-domain features, and the time-frequency-domain features are labeled with their respective physical meanings to obtain a training set;

[0014] Construct an SVM model, train the SVM model based on the training set, and obtain a diagnostic model;

[0015] Based on the actual vibration and acoustic emission signals of the blade under test during actual operation, the diagnostic model is used to diagnose the fault of the blade under test and obtain the diagnostic results.

[0016] The dynamic learning module is used to dynamically optimize the parameters of the diagnostic model.

[0017] The principle of this invention is as follows: A calibration module calibrates wind turbine blade specimens with normal, undamaged conditions and those with known typical damage, providing an initial benchmark for the system. The feature extraction module employs a multi-domain joint feature extraction method to extract time-domain, frequency-domain, and time-frequency-domain features, fusing these features to generate a high-dimensional fused feature vector. This comprehensively characterizes the signal properties, laying the foundation for subsequent effective classification. However, relying solely on manual calibration cannot accommodate individual differences in blades and unknown faults. Therefore, the wavelet clustering module performs unsupervised clustering analysis on the fused feature vector. Its purpose is to discover the potential distribution patterns of data points in the high-dimensional feature space. These patterns may exceed the pre-calibrated range, enabling the system to identify unknown, complex, or specific damage patterns under known typical damage (e.g., all belonging to the known category of fiber fracture, but with different degrees and directions of fracture, potentially leading to different cluster types). The system then assigns physical meaning to these newly discovered patterns, essentially combining prior knowledge with data-driven discovery. Subsequently, the fault diagnosis and early warning module uses the training set generated by clustering analysis to train a diagnostic model, achieving automatic fault diagnosis. The dynamic learning module ensures that this process is continuously optimized. Therefore, through the collaboration of multiple modules, this system overcomes the shortcomings of traditional methods that rely on fixed thresholds and cannot adapt to complex and variable operating conditions and individual differences, effectively improving the accuracy of wind turbine blade fault identification.

[0018] Furthermore, the clustering analysis of the fused feature vector based on the wavelet clustering algorithm includes:

[0019] Define the number of dimensions d of the fused feature vector, and the numerical range of the feature values ​​of any dimension;

[0020] Divide the numerical range of the feature values ​​of any one dimension into m segments, forming a sequence including m... d The feature space of each grid cell;

[0021] Based on the feature values ​​of the fused feature vector in each dimension, the fused feature vector is assigned to the corresponding grid cells in the feature space to obtain a spatial grid density distribution map;

[0022] A second wavelet transform and an inverse wavelet transform are performed on the spatial grid density distribution map to obtain an enhanced spatial grid density distribution map.

[0023] Calculate the local density and relative distance of any data point in the enhanced spatial grid density distribution map;

[0024] A decision graph is drawn based on the local density and the relative distance. Cluster centers are determined based on the decision graph, and cluster categories are obtained based on the cluster centers.

[0025] Directly clustering high-dimensional fused feature vectors can lead to excessively high dimensionality, computational complexity, and significant noise interference, making it difficult to accurately capture the data's intrinsic structure. This system addresses this by quantizing the feature space, discretizing the continuous, high-dimensional feature space into a low-dimensional density distribution map composed of grid cells, achieving efficient dimensionality reduction. However, simple quantization can lose detailed information. Therefore, this system treats the grid density distribution map as a signal, utilizing the powerful multi-scale analysis capabilities of wavelet transform to amplify the true dense regions (potential clusters) while filtering out pseudo-dense points caused by noise or quantization errors. The density map enhanced by wavelet transform exhibits significantly improved quality. Based on this, calculating local density and relative distance and drawing a decision map allows for the automatic and accurate identification of points with both high local density and high relative distance as cluster centers on a clearer data base. This overcomes the shortcomings of traditional clustering algorithms in high-dimensional, nonlinear data, such as the need to pre-determine the number of clusters and sensitivity to initial values ​​and noise.

[0026] Furthermore, the formula for the first wavelet transform is:

[0027] ;

[0028] in, Here, represents the wavelet transform coefficients, a is the scaling parameter, b is the translation parameter, and f(t) is the original acoustic emission signal. For the mother wavelet function, for The complex conjugate of , where t is time.

[0029] Acoustic emission signals are typical non-stationary transient signals, and their damage information exists simultaneously in the time and frequency dimensions. Traditional Fourier transforms can only provide global frequency domain characteristics of the signal. The introduction of the first wavelet transform utilizes an adaptive window of the wavelet basis function in the time-frequency plane. Through continuous changes in the scale parameter 'a' and the translation parameter 'b', it achieves time-frequency localization analysis of the signal, obtaining time-frequency characteristics. These time-frequency domain characteristics can characterize the properties of transient events such as crack propagation at different scales, solving the problem that traditional time-domain or frequency-domain characteristics cannot fully characterize non-stationary signals, and providing a foundation for accurately distinguishing different types of damage.

[0030] Furthermore, the formula for the second wavelet transform is:

[0031] ;

[0032] in, Here, represents the discrete wavelet transform coefficients, j is the decomposition scale parameter, k is the translation index, n is the time index of the discrete signal, and N is the total number of sampling points of the signal. For input signal, For the mother wavelet function, for .

[0033] Specifically, through a second wavelet transform using a fixed decomposition scale, the original spatial grid density map is decomposed into approximate and detail components of different frequency subbands. This multi-resolution decomposition effectively separates the macroscopic trends and local details / noise in the distribution map. During reconstruction using inverse wavelet transform, detail components beneficial for cluster identification can be selectively enhanced while noise components are suppressed, resulting in an enhanced density distribution map. This enhanced map essentially represents the core structural information of the original data after purification, enabling subsequent density peak calculations to be performed with a high signal-to-noise ratio, thus improving the accuracy and robustness of cluster center identification.

[0034] Furthermore, the formula for calculating the local density is as follows:

[0035] ;

[0036] in, Let be the local density of data point i, used to characterize the density of data points in the area surrounding data point i. Let i be the Euclidean distance between data point i and data point j. To cut off the distance.

[0037] Furthermore, the formula for calculating the relative distance is:

[0038] ;

[0039] in, Let be the relative distance to data point i, and let represent the minimum distance from data point i to any data point j with a higher density. Let be the local density of data point j.

[0040] If a data point has a high local density and a large relative distance, it means that it is not only the center of a local region, but also far away from other regions with higher density, which indicates that it is an independent cluster center.

[0041] Furthermore, the step of labeling the time-domain features, the frequency-domain features, and the time-frequency-domain features with their corresponding physical meanings to obtain a training set includes:

[0042] The time-domain features and the frequency-domain features are fused to obtain a vibration feature vector. The vibration feature vector and its corresponding physical meaning are labeled to obtain a vibration training set.

[0043] The time-frequency domain features are labeled with their corresponding physical meanings to obtain an acoustic emission training set.

[0044] In the training set construction process, this system processes the features of vibration signals and acoustic emission signals separately, and forms vibration training sets and acoustic emission training sets with their corresponding physical meaning labels. This separate processing method takes into account the different physical characteristics and representation capabilities of vibration and acoustic emission signals, and prevents information confusion that may be caused by direct fusion during the training stage. This allows the subsequently trained SVM model to learn the most discriminative patterns in each signal more accurately.

[0045] Furthermore, the construction of the SVM model, and the training of the SVM model based on the training set to obtain the diagnostic model, includes:

[0046] Construct the first SVM model and the second SVM model respectively;

[0047] The first SVM model is trained based on the vibration training set to obtain a vibration diagnostic model, and the second SVM model is trained based on the acoustic emission training set to obtain an acoustic emission diagnostic model.

[0048] The vibration diagnostic model and the acoustic emission diagnostic model can work in parallel, providing diagnostic opinions from two different dimensions. The advantages of this approach are: firstly, it improves system reliability; for example, when one sensor signal is interfered with, the other model can still provide valid information; secondly, when the diagnostic results of the two models are inconsistent, a more in-depth secondary verification can be triggered. This solves the problem of misjudgment easily produced by a single model due to limited or interfered input information.

[0049] Furthermore, based on the actual vibration signal and actual acoustic emission signal of the blade under test during actual field operation, the diagnostic model is used to diagnose the fault of the blade under test and obtain diagnostic results, including:

[0050] Extract the actual time-domain features and actual frequency-domain features of the actual vibration signal, fuse the actual time-domain features and actual frequency-domain features into an actual vibration feature vector, input the actual vibration feature vector into the vibration diagnosis model, and output the vibration diagnosis result;

[0051] Extract the actual time-frequency domain features of the actual acoustic emission signal, input the actual time-frequency domain features into the acoustic emission diagnostic model, and output the acoustic emission diagnostic results;

[0052] If the vibration diagnosis result is consistent with the acoustic emission diagnosis result, then the final diagnosis result is output;

[0053] If the vibration diagnosis result is inconsistent with the acoustic emission diagnosis result, a secondary verification process is initiated.

[0054] When the results are consistent, the diagnostic conclusions of the two models corroborate each other, indicating extremely high reliability, and the results are output directly. When the results are inconsistent, it indicates a complex situation (such as early damage, sensor failure, or strong interference). In this case, the system initiates a secondary verification process for further analysis.

[0055] Furthermore, the inconsistency between the vibration diagnosis results and the acoustic emission diagnosis results includes:

[0056] The vibration diagnosis result is normal, and the acoustic emission diagnosis result is abnormal.

[0057] The secondary verification process includes:

[0058] Obtain the root mean square value of the actual vibration signal and generate a root mean square value time series curve;

[0059] The rotation period of the blade is obtained by analyzing the root mean square value time series curve.

[0060] Extract the event rate per second of the actual acoustic emission signal and generate an acoustic emission event rate time-series curve;

[0061] Align the root mean square value time series curve with the acoustic emission event rate time series curve, and cut the curve with the rotation period as the length.

[0062] Observe the time-series curve of the acoustic emission event rate after cutting. If the acoustic emission event rate fluctuates synchronously with the rotation period, the acoustic emission diagnosis result shall be used as the final diagnosis result; otherwise, the vibration diagnosis result shall be used as the final diagnosis result.

[0063] The principle behind this verification process utilizes the periodic physical law of blade rotation. Due to the influence of gravity, cracks on the blade open during the lower half of the rotation (under tension) and close during the upper half of the rotation (under compression), resulting in a modulation phenomenon in which the acoustic emission event rate is synchronized with the rotation period. Therefore, verifying this phenomenon can elevate an isolated acoustic emission event alarm to a damage diagnosis supported by physical laws, achieving secondary verification.

[0064] One or more technical solutions provided by this invention have at least the following technical effects or advantages:

[0065] 1. This invention establishes a complete technology chain from signal perception, feature fusion, intelligent clustering to decision diagnosis. Its advantage lies in using wavelet clustering algorithm to perform unsupervised learning on the fused feature vectors, automatically discovering damage patterns, and overcoming the drawbacks of traditional methods that rely on preset thresholds and fixed damage templates. By constructing training sets and SVM models for vibration and acoustic emission signals respectively, it achieves targeted diagnosis of macroscopic mechanical states and microscopic damage events, solving the problem of insufficient monitoring dimensions and susceptibility to misjudgment by single technical means.

[0066] 2. This system transforms high-dimensional, nonlinear feature distributions into clear clustering structures through feature space quantization and wavelet transform enhancement. Then, by calculating local density and relative distance, it automatically and accurately identifies cluster centers. This process overcomes the shortcomings of traditional clustering algorithms when processing blade monitoring data, such as sensitivity to noise, the need to preset the number of categories, and difficulty in identifying non-spherical clusters, resulting in more precise differentiation of damage patterns.

[0067] 3. When the vibration and acoustic emission diagnostic results are inconsistent, this system does not simply discard the results, but instead initiates a physical verification process based on the periodicity of blade rotation. It uses the regular influence of gravity on cracks to verify the authenticity of acoustic emission events, which greatly reduces the false alarm rate.

[0068] 4. Through the dynamic learning module, the system can continuously optimize the diagnostic model parameters based on new monitoring data, enabling the system to adapt to individual differences of different blades (such as manufacturing tolerances and installation positions) and performance degradation during long-term operation, and establish a unique diagnostic model for each blade, thereby overcoming the problem that the general model is not adapted to individual differences. Attached Figure Description

[0069] The accompanying drawings, which are provided to further illustrate embodiments of the invention and constitute a part of this invention, are not intended to limit the scope of the invention.

[0070] Figure 1 This is a schematic diagram of the composition of a wind turbine blade condition monitoring and fault early warning system based on wavelet clustering algorithm in this invention. Detailed Implementation

[0071] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, where there is no conflict, the embodiments of the present invention and the features thereof can be combined with each other.

[0072] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0073] Example 1

[0074] Please refer to Figure 1 Embodiment 1 of the present invention provides a wind turbine blade condition monitoring and fault early warning system based on wavelet clustering algorithm, the system comprising:

[0075] The calibration module is used to calibrate wind turbine blade specimens that are normal and undamaged as well as those with a known type of typical damage, and to obtain calibration results. The known typical damage includes type M.

[0076] The data acquisition module is used to acquire the original vibration signals and original acoustic emission signals of the wind turbine blade specimen under simulated operating conditions;

[0077] The feature extraction module is used to extract the time-domain and frequency-domain features of the original vibration signal, and to extract the time-frequency domain features of the original acoustic emission signal based on the first wavelet transform, and to fuse the time-domain features, the frequency-domain features and the time-frequency domain features to generate a fused feature vector;

[0078] The wavelet clustering module is used to perform cluster analysis on the fused feature vector based on the wavelet clustering algorithm to obtain N cluster categories, and assign a physical meaning to any cluster category based on the calibration result;

[0079] The fault diagnosis and early warning module is used for:

[0080] The time-domain features, the frequency-domain features, and the time-frequency-domain features are labeled with their respective physical meanings to obtain a training set;

[0081] Construct an SVM model, train the SVM model based on the training set, and obtain a diagnostic model;

[0082] Based on the actual vibration and acoustic emission signals of the blade under test during actual operation, the diagnostic model is used to diagnose the fault of the blade under test and obtain the diagnostic results.

[0083] The dynamic learning module is used to dynamically optimize the parameters of the diagnostic model.

[0084] In practice, the calibration module first performs offline calibration. Fiberglass composite blade specimens are selected, and typical damages, such as matrix cracking, fiber breakage, and delamination damage, are artificially created in a laboratory environment. Multiple specimens are created for each type of damage. These damage types are then calibrated together with undamaged normal specimens to obtain the calibration results.

[0085] The data acquisition module uses a low-frequency accelerometer to collect the original vibration signal and a resonant broadband sensor to collect the original acoustic emission signal. It also uses an adaptive filtering algorithm (using an LMS adaptive filter) to remove environmental noise interference.

[0086] The time-domain features include: amplitude; rise time, which is the time it takes for the signal to reach its maximum peak value from a threshold value; duration, which is the total time the signal exceeds the threshold value; ring count, which is the peak count within the duration; and energy, which is the area enclosed by the signal envelope and the time axis. The frequency-domain features include: peak frequency; centroid frequency, which is the centroid frequency of the power spectrum; power spectral density, which is the energy distribution of the signal in the frequency domain; and frequency standard deviation, which is the degree of dispersion of the frequency distribution. The time-frequency-domain features include: wavelet energy spectrum, which is the sum of the squares of the wavelet coefficients; wavelet entropy, which is the entropy value of the wavelet energy distribution; and wavelet scale-amplitude features, which are the amplitude distributions at different scales.

[0087] In this embodiment, the wavelet clustering module can automatically detect damage patterns that exceed the pre-calibrated range using the wavelet clustering algorithm. For example, the system identifies a new cluster category from the feature space, with characteristic parameters of an amplitude range of 55.2~78.3 dB and a peak frequency range of 70~110 kHz (between matrix cracking and fiber fracture). Field inspection revealed that this category corresponds to a composite damage mode of "matrix-fiber interface debonding," a new mode not covered during the calibration phase.

[0088] In the fault diagnosis and early warning module, the SVM model uses a radial basis kernel function, and the specific model training method is an existing technology in this field.

[0089] The dynamic learning module optimizes parameters through an online learning engine. When the system experiences three consecutive false alarms or missed alarms, the learning process is automatically triggered. The system also periodically initiates a global optimization process during off-peak hours (such as low wind speed periods), using all accumulated new data to comprehensively retrain the model. Specifically, incremental learning algorithms are used to update SVM model parameters (e.g., adjusting the penalty coefficient C from 1.0 to 0.8), and a rolling update strategy is employed. The new model is deployed after testing and validation in a sandbox environment to ensure system stability. The system can also establish personalized health benchmarks for each blade, such as updating the normal range benchmark for vibration amplitude using a moving average method to adapt to normal blade aging.

[0090] The step of performing cluster analysis on the fused feature vector based on the wavelet clustering algorithm includes:

[0091] Define the number of dimensions d of the fused feature vector, and the numerical range of the feature values ​​of any dimension;

[0092] Divide the numerical range of the feature values ​​of any one dimension into m segments, forming a sequence including m... d The feature space of each grid cell;

[0093] Based on the feature values ​​of the fused feature vector in each dimension, the fused feature vector is assigned to the corresponding grid cells in the feature space to obtain a spatial grid density distribution map;

[0094] A second wavelet transform and an inverse wavelet transform are performed on the spatial grid density distribution map to obtain an enhanced spatial grid density distribution map.

[0095] Calculate the local density and relative distance of any data point in the enhanced spatial grid density distribution map;

[0096] A decision graph is drawn based on the local density and the relative distance. Cluster centers are determined based on the decision graph, and cluster categories are obtained based on the cluster centers.

[0097] In determining the number of dimensions d of the fused feature vector, principal component analysis (PCA) can be used to screen features in order to reduce computational complexity and improve clustering effect. For example, in acoustic emission signal analysis, PCA screens out four key features: amplitude, RA value, peak frequency and centroid frequency, which significantly reduces the dimensionality.

[0098] When dividing a numerical range into segments, a heuristic method can be used to find the optimal quantization result. First, set m to a large value (overquantization) and try to find clusters; if no suitable clusters are found, gradually decrease the value of m and repeat the above process until a suitable cluster is found.

[0099] The purpose of performing a second wavelet transform and inverse wavelet transform on the spatial grid density distribution map is to enhance dense regions in the feature space, suppress noise, and thus optimize the clustering effect. In practice, the spatial grid density distribution map is treated as a two-dimensional discrete signal, with the density value of each grid cell used as the signal amplitude. The density value is defined as the number of feature vectors contained within that grid cell; for example, if a grid cell is assigned 15 feature vectors, its density value is 15. Finally, a multi-scale wavelet coefficient matrix is ​​obtained, including approximation coefficients and detail coefficients. Thresholding is applied to the wavelet coefficients to enhance the signal, i.e., retaining significant coefficients (such as coefficients with an absolute value greater than a threshold T) and suppressing noise coefficients (such as coefficients less than T that are set to zero). The threshold T can be determined by those skilled in the art based on the actual situation.

[0100] Subsequently, the processed coefficients are reconstructed using inverse wavelet transform to obtain the enhanced spatial grid density distribution map. The specific transformation formula is existing technology, and those skilled in the art can select it as needed.

[0101] When drawing the decision graph, a scatter plot is created with local density on the horizontal axis and relative distance on the vertical axis. Cluster centers are located in the upper right corner of the decision graph, representing data points with both high local density and high relative distance. For example, if a data point has a local density greater than a density threshold and a relative distance greater than a distance threshold, then that data point is identified as a cluster center.

[0102] The specific density and distance thresholds can be determined by those skilled in the art in conjunction with the actual situation. Once the cluster centers are identified, the system then uses connectivity analysis (such as 4-connection) to merge connected salient grid cells into a single cluster category based on the cluster centers, and finally outputs the cluster labels.

[0103] The formula for the first wavelet transform is:

[0104] ;

[0105] in, Here, represents the wavelet transform coefficients, a is the scaling parameter, b is the translation parameter, and f(t) is the original acoustic emission signal. For the mother wavelet function, for The complex conjugate of , where t is time.

[0106] The values ​​of the scale and translation parameters need to be determined by those skilled in the art based on the actual situation. The mother wavelet function can be either the Morlet wavelet or the Mexican hat wavelet. After transformation, the wavelet energy spectrum and wavelet entropy are extracted. The wavelet energy spectrum is obtained by calculating the sum of squares of each scale coefficient, and the wavelet entropy is calculated based on the energy distribution.

[0107] The formula for the second wavelet transform is:

[0108] ;

[0109] in, Here, represents the discrete wavelet transform coefficients, j is the decomposition scale parameter, k is the translation index, n is the time index of the discrete signal, and N is the total number of sampling points of the signal. For input signal, For the mother wavelet function, for The complex conjugate of . The decomposition scale parameter j is set to 4, and the mother wavelet function is the Daubechies wavelet (db2) function.

[0110] The formula for calculating the local density is as follows:

[0111] ;

[0112] in, Let be the local density of data point i, used to characterize the density of data points in the area surrounding data point i. Let be the Euclidean distance between data point i and data point j. The weights are set to the coefficients of variation for each feature dimension; for example, the coefficient of variation for amplitude is 0.85, and the coefficient of variation for peak frequency is 0.62. The cutoff distance is determined through information entropy optimization, for example, by calculating different... The value corresponding to the clustering result information entropy is selected to maximize the information entropy.

[0113] The formula for calculating the relative distance is as follows:

[0114] ;

[0115] in, Let be the relative distance to data point i, and let represent the minimum distance from data point i to any data point j with a higher density. Let be the local density of data point j.

[0116] The step of labeling the time-domain features, the frequency-domain features, and the time-frequency-domain features with their corresponding physical meanings to obtain a training set includes:

[0117] The time-domain features and the frequency-domain features are fused to obtain a vibration feature vector. The vibration feature vector and its corresponding physical meaning are labeled to obtain a vibration training set.

[0118] The time-frequency domain features are labeled with their corresponding physical meanings to obtain an acoustic emission training set.

[0119] In practice, the fusion of time-domain features and frequency-domain features can be achieved by direct splicing, weighted fusion, or dimensionality reduction through principal component analysis. Those skilled in the art can choose the appropriate method based on the specific circumstances.

[0120] The step of constructing an SVM model, training the SVM model based on the training set to obtain a diagnostic model, includes:

[0121] Construct the first SVM model and the second SVM model respectively;

[0122] The first SVM model is trained based on the vibration training set to obtain a vibration diagnostic model, and the second SVM model is trained based on the acoustic emission training set to obtain an acoustic emission diagnostic model.

[0123] The diagnostic model is used to diagnose the faults of the blade under test based on the actual vibration and acoustic emission signals from actual on-site operation, and to obtain diagnostic results, including:

[0124] Extract the actual time-domain features and actual frequency-domain features of the actual vibration signal, fuse the actual time-domain features and actual frequency-domain features into an actual vibration feature vector, input the actual vibration feature vector into the vibration diagnosis model, and output the vibration diagnosis result;

[0125] Extract the actual time-frequency domain features of the actual acoustic emission signal, input the actual time-frequency domain features into the acoustic emission diagnostic model, and output the acoustic emission diagnostic results;

[0126] If the vibration diagnosis result is consistent with the acoustic emission diagnosis result, then the final diagnosis result is output;

[0127] If the vibration diagnosis result is inconsistent with the acoustic emission diagnosis result, a secondary verification process is initiated.

[0128] The extraction methods for actual time-domain features, actual frequency-domain features, and actual time-frequency-domain features are consistent with those used in the previous training of the diagnostic model.

[0129] The inconsistency between the vibration diagnosis result and the acoustic emission diagnosis result includes:

[0130] The vibration diagnosis result is normal, and the acoustic emission diagnosis result is abnormal.

[0131] The secondary verification process includes:

[0132] Obtain the root mean square value of the actual vibration signal and generate a root mean square value time series curve;

[0133] The rotation period of the blade is obtained by analyzing the root mean square value time series curve.

[0134] Extract the event rate per second of the actual acoustic emission signal and generate an acoustic emission event rate time-series curve;

[0135] Align the root mean square value time series curve with the acoustic emission event rate time series curve, and cut the curve with the rotation period as the length.

[0136] Observe the time-series curve of the acoustic emission event rate after cutting. If the acoustic emission event rate fluctuates synchronously with the rotation period, the acoustic emission diagnosis result shall be used as the final diagnosis result; otherwise, the vibration diagnosis result shall be used as the final diagnosis result.

[0137] The rotation period is obtained through spectral analysis of the time-series curve. The acoustic emission event rate refers to the number of acoustic emission events occurring per unit time. The acoustic emission event rate exhibits fluctuations synchronously with the rotation period, meaning that the acoustic emission event rate shows regular changes within each period. For example, assuming the blade's rotation period is 4 seconds, the event rate increases significantly during the second half of the period (0-2 seconds), while it decreases significantly during the second half of the period (2-4 seconds).

[0138] Furthermore, the inconsistency between the vibration diagnosis result and the acoustic emission diagnosis result also includes situations where the vibration diagnosis result is abnormal and the acoustic emission diagnosis result is normal, or both are abnormal but the diagnosed fault types are inconsistent.

[0139] The former system uses acoustic emission diagnostic results as the final diagnostic result and marks "microscopic damage warning" in the output; the latter system prompts "unknown complex fault" and requests manual intervention. After manual confirmation of the result, the data is returned to the dynamic learning module for continuous model learning.

[0140] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0141] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A wind turbine blade condition monitoring and fault warning system based on wavelet clustering algorithm, characterized in that, The system comprises: a calibration module configured to calibrate wind turbine blade test pieces without damage and with a known typical damage, and obtain calibration results, wherein the known typical damage comprises M types; a data acquisition module configured to acquire original vibration signals and original acoustic emission signals of the wind turbine blade test pieces in a simulated operating environment; a feature extraction module configured to extract time domain features and frequency domain features of the original vibration signals, and extract time-frequency domain features of the original acoustic emission signals based on a first wavelet transform, and fuse the time domain features, the frequency domain features and the time-frequency domain features to generate a fusion feature vector; a wavelet clustering module configured to perform clustering analysis on the fusion feature vector based on a wavelet clustering algorithm, obtain N clustering categories, and assign a physical meaning to any clustering category based on the calibration results; a fault diagnosis and early warning module configured to: label the time domain features, the frequency domain features and the time-frequency domain features with respective corresponding physical meanings to obtain a training set; construct an SVM model, train the SVM model based on the training set, and obtain a diagnosis model; perform fault diagnosis on a to-be-detected blade based on actual vibration signals and actual acoustic emission signals of the to-be-detected blade in actual operation, and obtain a diagnosis result; a dynamic learning module configured to dynamically optimize parameters of the diagnosis model.

2. The wind turbine blade condition monitoring and fault warning system based on wavelet clustering algorithm according to claim 1, characterized in that, The wavelet clustering algorithm-based clustering analysis on the fusion feature vector comprises: defining a dimension number d of the fusion feature vector and a numerical range of feature values of any one dimension; The numerical range of the eigenvalue of any one dimension is divided into m sections to form a feature space including m d grid cells. based on the feature values of the fusion feature vector in each dimension, distributing the fusion feature vector in corresponding grid cells of the feature space to obtain a spatial grid density distribution map; performing a second wavelet transform and a wavelet inverse transform on the spatial grid density distribution map to obtain an enhanced spatial grid density distribution map; calculating a local density and a relative distance of any data point in the enhanced spatial grid density distribution map; drawing a decision graph based on the local density and the relative distance, determining a clustering center based on the decision graph, and obtaining a clustering category based on the clustering center. 3.The wind turbine blade condition monitoring and fault warning system based on wavelet clustering algorithm of claim 1, wherein, The formula of the first wavelet transform is: ； wherein is a wavelet transform coefficient, a is a scale parameter, b is a translation parameter, f(t) is the original acoustic emission signal, is a mother wavelet function, is the complex conjugate of, t is time.

4. The wind turbine blade condition monitoring and fault warning system based on wavelet clustering algorithm according to claim 2, characterized in that, The formula of the second wavelet transform is: ； wherein, is a discrete wavelet transform coefficient, j is a decomposition scale parameter, k is a translation index, n is a time index of the discrete signal, N is the total number of samples of the signal, is an input signal, is a mother wavelet function, is a complex conjugate of .

5. The wind turbine blade condition monitoring and fault warning system based on wavelet clustering algorithm according to claim 2, characterized in that, The calculation formula of the local density is: ； wherein, is the local density of data point i, used to characterize the data point density of the region around data point i, is the Euclidean distance between data point i and data point j, is the cut-off distance.

6. The wind turbine blade condition monitoring and fault warning system based on wavelet clustering algorithm according to claim 5, characterized in that, The calculation formula of the relative distance is: ； wherein, is the relative distance of data point i, denoting the minimum distance of data point i to any data point j with higher density than it, is the local density of data point j.

7. The wind turbine blade condition monitoring and fault warning system based on wavelet clustering algorithm according to claim 1, characterized in that, The labeling of the time domain features, the frequency domain features and the time-frequency domain features with respective corresponding physical meanings to obtain the training set comprises: fusing the time domain features and the frequency domain features to obtain a vibration feature vector, and labeling the vibration feature vector with a corresponding physical meaning to obtain a vibration training set; labeling the time-frequency domain features with a corresponding physical meaning to obtain an acoustic emission training set.

8. The wind turbine blade condition monitoring and fault warning system based on wavelet clustering algorithm according to claim 7, characterized in that, The construction of the SVM model, the training of the SVM model based on the training set, and the obtaining of the diagnosis model comprise: constructing a first SVM model and a second SVM model respectively; training the first SVM model based on the vibration training set to obtain a vibration diagnosis model, and training the second SVM model based on the acoustic emission training set to obtain an acoustic emission diagnosis model. 9.The wind turbine blade condition monitoring and fault warning system based on wavelet clustering algorithm of claim 8, wherein, The actual vibration signal and the actual acoustic emission signal of the to-be-detected blade based on the actual operation are used to diagnose the fault of the to-be-detected blade by using the diagnosis model, and a diagnosis result is obtained, including: extracting actual time domain features and actual frequency domain features of the actual vibration signal, fusing the actual time domain features and the actual frequency domain features into an actual vibration feature vector, inputting the actual vibration feature vector into the vibration diagnosis model, and outputting a vibration diagnosis result; extracting actual time-frequency domain features of the actual acoustic emission signal, inputting the actual time-frequency domain features into the acoustic emission diagnosis model, and outputting an acoustic emission diagnosis result; if the vibration diagnosis result is consistent with the acoustic emission diagnosis result, outputting a final diagnosis result; if the vibration diagnosis result is inconsistent with the acoustic emission diagnosis result, starting a secondary verification process.

10. The wind turbine blade condition monitoring and fault warning system based on wavelet clustering algorithm according to claim 9, characterized in that, The more inconsistent the vibration diagnosis result is with the acoustic emission diagnosis result, including: the vibration diagnosis result is normal, and the acoustic emission diagnosis result is abnormal; the secondary verification process includes: obtaining a root mean square value of the actual vibration signal, and generating a root mean square value time sequence curve; analyzing the root mean square value time sequence curve to obtain a rotation period of the blade; extracting an event rate per second of the actual acoustic emission signal, and generating an acoustic emission event rate time sequence curve; aligning the root mean square value time sequence curve with the acoustic emission event rate time sequence curve, and cutting with the rotation period as the length; observing the acoustic emission event rate time sequence curve after cutting, if the acoustic emission event rate presents a fluctuation condition synchronized with the rotation period, taking the acoustic emission diagnosis result as the final diagnosis result, otherwise, taking the vibration diagnosis result as the final diagnosis result.

Images