Aerodynamic noise analysis-based online health state monitoring method for blades of wind driven generator
By using aerodynamic noise analysis and the DBSCAN clustering algorithm, non-invasive real-time health status monitoring of wind turbine blades was achieved, solving the problem of real-time detection in existing technologies and improving the recall rate and generalization performance of the model.
Patent Information
- Application Number
- CN202411106922.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-13
- Publication Date
- 2024-11-12
AI Technical Summary
Existing technologies cannot effectively reduce the dimensionality of high-frequency dense sampling signals to capture key features, resulting in the curse of dimensionality in similarity calculation and cluster analysis. They are unable to perform threat assessment and alarm calculation, and cannot achieve data sampling, feature extraction and anomaly detection while the wind turbine is running continuously, lacking real-time monitoring and early warning.
By collecting aerodynamic noise signals from wind turbine blades and generating energy spectrum vectors, online adaptive diagnosis using the DBSCAN clustering algorithm is performed through unsupervised learning. Combined with Euclidean distance and mechanistic criteria, abnormal operating conditions are warned, achieving non-intrusive real-time monitoring.
It realizes non-invasive real-time health status monitoring of wind turbine blades, and can adaptively detect abnormal operating conditions in real time and issue early warnings even in the absence of abnormal sample data, thereby improving the recall rate of detection and the generalization performance of the model.
Smart Images

Figure CN118934493A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a technology in the field of wind power generation, in particular to an online health status monitoring method for wind turbine blades based on aerodynamic noise analysis. Background Art
[0002] The current wind turbine blade fault detection technology is mainly based on data-driven methods. However, the diagnostic ability of the model built by the data-driven method depends on the collection of a large number of wind turbine blade samples, which has high data costs and a small proportion of abnormal samples, resulting in a low recall rate of the detection model. In addition, data-driven methods are mostly used for fault detection rather than real-time health status monitoring under online operation. The wind turbine blade fault detection technology based on pure mechanism drive is adapted to the complex working conditions of the wind turbine during operation, resulting in poor robustness of the detection model and a lack of historical data to introduce into the model to enhance the generalization performance. Summary of the invention
[0003] The present invention aims to solve the problems that the prior art cannot effectively reduce the dimension of high-frequency densely sampled signals to capture the key features of high-frequency densely sampled signals, resulting in dimensional disasters in similarity calculation and cluster analysis, and cannot perform threat assessment and alarm calculation. In addition, the present invention cannot realize data sampling, feature extraction, anomaly detection and operating condition warning simultaneously when the wind turbine is running without stopping, and cannot realize non-invasive real-time monitoring and early warning of wind turbines. A wind turbine blade health status monitoring method based on aerodynamic noise analysis is proposed, which can improve the health status detection performance of wind turbine blades, enhance the model generalization performance through historical data while detecting the aerodynamic noise mechanism, and can perform online adaptive wind turbine blade diagnosis and model adjustment.
[0004] The present invention is achieved through the following technical solutions:
[0005] The present invention relates to a method for monitoring the health status of wind turbine blades based on aerodynamic noise analysis. The aerodynamic noise signal of each blade is collected and a corresponding energy spectrum vector is generated, which is then merged with energy spectrum vector data under historical normal working conditions to obtain an energy spectrum vector cluster. For the energy spectrum vector cluster of each blade, an online adaptive diagnosis based on unsupervised learning is performed using a DBSCAN clustering algorithm to identify abnormal noise points and provide abnormal working condition warnings.
[0006] Preferably, when the similarity between the energy spectrum vector of the blade and the energy spectrum vector under historical normal operating conditions exceeds a threshold, an abnormal operating condition warning is issued.
[0007] The similarity refers to: the sum of the Euclidean distances of the energy spectrum vector of the current leaf relative to other energy spectrum vectors.
[0008] The energy spectrum vector is obtained in the following way:
[0009] a) performing analog-to-digital conversion on the collected aerodynamic noise signal of each blade and further performing low-pass digital filtering;
[0010] b) performing adaptive signal variable mode decomposition on the blade aerodynamic noise signal obtained in step a;
[0011] c) extracting the modal functions of the first several scales of the signal after decomposition in step b, calculating the energy of each modal function and the energy percentage representation of each modal function, and then obtaining an energy spectrum vector composed of the energy percentage representation. Technical Effects
[0012] The present invention obtains the characteristic vector corresponding to the aerodynamic noise of the wind turbine blades through noise reduction preprocessing, adaptive variable mode decomposition and energy spectrum calculation, and realizes real-time monitoring of the health status through online sampling, feature extraction, mechanistic criteria and unsupervised cluster analysis. Compared with the prior art, the present invention can perform non-invasive real-time monitoring of the health status of the wind turbine blades while the wind turbine remains in operation without shutting down; in the absence of a detection model driven by abnormal sample training data, it can perform adaptive real-time detection of abnormal working conditions of the wind turbine without relying on abnormal sample data, and perform abnormal working condition warning. BRIEF DESCRIPTION OF THE DRAWINGS
[0013] Figure 1 It is a schematic diagram of the system of the present invention;
[0014] Figure 2 It is a flow chart of the present invention;
[0015] Figure 3 This is a schematic diagram of the communication layer architecture of an embodiment;
[0016] Figure 4 This is an example effect diagram. DETAILED DESCRIPTION
[0017] like Figure 1As shown, an online health status monitoring system for wind turbine blades based on aerodynamic noise analysis involved in this embodiment includes: a sampling unit, a feature extraction unit, a detection unit and an early warning unit, wherein: the sampling unit performs low-pass filtering noise reduction preprocessing according to the aerodynamic noise audio information of the wind turbine blades to obtain the aerodynamic noise audio signal after noise reduction; the feature extraction unit performs adaptive variable mode decomposition and energy spectrum calculation according to the aerodynamic noise audio signal after noise reduction to extract the energy spectrum vector corresponding to the aerodynamic noise audio signal; the detection unit performs similarity judgment, Euclidean distance calculation and DBSCAN cluster analysis according to the energy spectrum vectors corresponding to a series of aerodynamic noise audio signals obtained by sampling and feature extraction; the early warning unit issues an early warning of the corresponding blade health status according to the cluster analysis result and the similarity and Euclidean distance criteria.
[0018] like Figure 2 As shown, a method for online health status monitoring of wind turbine blades based on the above system includes:
[0019] Step 1: Deploy the sensor units on the fan blades: For each fan, three sensor units are required to sample the aerodynamic noise signals when the blades are running. Since the aerodynamic noise generated when the fan blades are running is mainly distributed at the end of the blades, the sensors should be deployed at a suitable position near the end of the blades.
[0020] Step 2: Filter out high-frequency noise from the data collected in step 1, specifically including:
[0021] 2.1 Construct a low-pass filter whose transfer function is: Where: s is the differential operator, ω c is the filter cutoff frequency. The cutoff frequency will be constructed based on the noise frequency band of the field sampling signal. c different.
[0022] 2.2 The low-pass filter transfer function is transformed through Tuspin to obtain a discretized differential equation, and the collected data is converted into digital form and then further low-pass digitally filtered.
[0023] Step 3, feature decomposition: performing adaptive signal variable mode decomposition on the blade aerodynamic noise signal X(t) obtained in step 2, specifically including:
[0024] 3.1 Find the maximum and minimum points of X(t), and use cubic spline interpolation to fit the envelope u(t) and v(t);
[0025] 3.2 Subtract its own mean envelope from X(t), specifically:
[0026] 3.3 Determine whether the number of extreme points and zero-crossing points in h(t) satisfies |ab|≤1 and whether the two envelopes of h(t) are symmetrical based on the horizontal coordinate. If so, take h(t) as the first mode function IMF1(t). Otherwise, let X(t) = h(t) and repeat step 2, where: a is the number of extreme points in h(t), and b is the number of zero-crossing points in h(t);
[0027] 3.4 Subtract the first modal function IMF1(t) from X(t) to obtain the first-order residual signal r1(t), that is, r1(t) = X(t) - IMF1(t). Take r1(t) as the new signal to be processed and repeat steps 3.1 to 3.3 to obtain a series of modal functions IMF1(t), IMF2(t), ...IMF n (t) and the residual signals r1(t),r2(t),...r n (t). When the IMF n (t) The corresponding residual signal r n (t) is less than the threshold value, the adaptive variable mode decomposition process ends.
[0028] 3.5 Through adaptive variable mode decomposition, the original blade aerodynamic noise signal is represented as a series of mode functions IMF1(t), IMF2(t), ...IMF n (t) and the residual signals r1(t),r2(t),...r n The sum of (t) is
[0029] Step 4: Extract features of the decomposed signal: extract the modal functions IMF1(t)~IMF1(t) of the first N scales of the decomposed signal. N (t), extract the energy spectrum vector as a feature. The energy of each modal function is obtained by using the discrete signal energy formula and Parseval theorem The energy percentage of each mode function is expressed as and the energy spectrum vector H p =[E p1 ,E p2 ,...,E pN ] T , where: T is the total data point length of mode function i.
[0030] The energy percentage representation, that is, the energy percentage of each modal function, represents the composition of modal components with different characteristics in the blade aerodynamic noise signal, and can effectively represent and distinguish blade audio signals with different aerodynamic noise characteristics.
[0031] Step 5: Collect the historical aerodynamic noise signals X of the three blades a, b, and c of each wind turbine under normal working conditions.a1 (t),X a2 (t),...,X am (t), X b1 (t),X b2 (t),...,X bm (t), X c1 (t),X c2 (t),...,X cm (t), perform signal processing from step 2 to step 4 on all historical aerodynamic noise signals to obtain a series of energy spectrum vectors H pa1 ,H pa2 ,...,H pam , H pb1 ,H pb2 ,...,H pbm , H pc1 ,H pc2 ,...,H pcm As data augmentation samples.
[0032] Step 6: Under online operation conditions, for each wind turbine's three-blade aerodynamic noise signal x a (t),x b (t),x c (t) Perform cross-correlation check, specifically: when any blade has a cross-correlation R with the other two blades xy If the value is smaller, an abnormal condition warning is issued for the blade, otherwise the next step is executed; specifically: Where: x, y are the aerodynamic noise signals of any two blades of the same wind turbine;
[0033] In this embodiment, when the difference between the cross-correlation of any blade and the other two blades is less than 0.01, an abnormal condition warning is issued for the blade.
[0034] Step 7: The blade aerodynamic noise signal x of the same wind turbine a (t),x b (t),x c (t) Repeat steps 2 to 4 to obtain the energy spectrum vector h of the aerodynamic noise signal of each blade. a ,h b ,h c Then, it is combined with the energy spectrum vector data under historical normal working conditions to obtain the energy spectrum vector cluster: H a =[H pa1 ,H pa2 ,...,H pam ,h a ] T =[H' pa1 ,H' pa2 ,...,H'pam ,H' pa(m+1) ] T , H b =[H pb1 ,H pb2 ,...,H pbm ,h b ] T =[H' pb1 ,H' pb2 ,...,H' pbm ,H' pb(m+1) ] T , H c =[H pc1 ,H pc2 ,...,H pcm ,h c ] T =[H' pc1 ,H' pc2 ,...,H' pcm ,H' pc(m+1) ] T .
[0035] Step 8: Determine the similarity between the new sampled data and the old data based on the vector Euclidean distance: For each energy spectrum vector cluster of each blade, calculate the sum of the Euclidean distances of each energy spectrum vector relative to other energy spectrum vectors.
[0036] For example, for the vector H' in the energy spectrum vector cluster of leaf a pai , its Euclidean distance relative to other energy spectrum vectors and Wherein: m+1 is the length of the energy spectrum vector cluster, j is the index of the element of each energy spectrum vector, and N is the length of each energy spectrum vector, which is equal to the number of modal function energy percentages contained in each energy spectrum vector.
[0037] Step 9: For each blade, find the energy spectrum vector with the largest Euclidean distance relative to other energy spectrum vectors. Taking blade a as an example, when the energy spectrum vector is the latest sampled energy spectrum vector h a When , go to step 10, otherwise go to step 11.
[0038] Step 10: Diagnose abnormal working conditions based on the Euclidean distance criterion: Taking blade a as an example, Eu a(m+1) ≥T a , when the condition is met, an abnormal operating condition warning is issued to blade a, otherwise, the process goes to step 11;
[0039] Step 11: For each blade energy spectrum vector cluster, apply the DBSCAN clustering algorithm to perform online adaptive diagnosis based on unsupervised learning to identify abnormal noise points: Taking blade a as an example, when the abnormal point is the energy spectrum vector h a, then an abnormal operating condition warning will be issued to the wind turbine generator and wind turbine blade a where blade a is located.
[0040] The DBSCAN clustering algorithm divides data points into core points, boundary points and abnormal noise points, where a core point means that there are at least MinPts other points in its ε-neighborhood, that is, N ε (p) = {q∈H a |d(p,q)≤ε}, where: d(p,q) is the distance between vector p and vector q. The condition for point p to be marked as a core point is |N ε (p)|≥MinPts; the boundary point refers to: when N ε If the number of points contained in (p) is less than MinPts, but p is located in the ε-neighborhood of other core points, then p is a boundary point, otherwise it is an abnormal noise point.
[0041] The abnormal operating condition warning refers to: sending an alarm command signal for blade a to the wind turbine blade online health status monitoring system.
[0042] Step 12: When the three blades are all detected as being in normal operating condition after the above steps, the wind turbine is operating normally, and a monitoring cycle is completed.
[0043] like Figure 3 As shown in the figure, after specific actual experiments, the above method is used to obtain the aerodynamic noise sampling signal data of a wind turbine generator set in Rudong wind farm, Jiangsu Province. Figure 4 As shown in the figure, the aerodynamic noise signals of the normal fan and the abnormal fan samples in this embodiment and their sampling and feature extraction results are shown. As can be seen from the figure, there are obvious differences between the feature vectors of the normal aerodynamic noise signal and the feature vectors of the abnormal aerodynamic noise signal in each component of the energy spectrum.
[0044] Secondly, the present invention is further evaluated on the MIMII dataset (Harsh Purohit et al., "MIMIIDataset: Sound Datasetfor Malfunctioning Industrial Machine Investigation and Inspection", 2019.): the dataset of each machine contains about 1,000 normal sound samples for training, and 100-200 samples of normal and abnormal sounds for testing. The distribution of normal and abnormal samples is shown in Table 1.
[0045] Table 1
[0046] Table 2
[0047] The detection accuracy is compared with other supervised learning-based methods. The results are shown in Table 2. The comparison indicators include precision Pr, recall rate Re and F1Score, which are as follows: Where: TP is the number of true positive samples, FP is the number of false positive samples, and FN is the number of false negative samples. The compared models include support vector classification, lightGBM, decision tree, random forest, K nearest neighbor, Gaussian naive Bayes and multilayer perceptron. The present invention achieves higher recall and F1 scores than supervised learning methods. Due to the imbalance of the data set, the recall and F1 scores of the SVM, KNN and MLP models are almost 0. The decision tree, random forest and lightGBM models also failed to achieve good indicators.
[0048] Compared with the prior art, the present invention not only achieved a classification accuracy of 75%, but also achieved a recall rate of 0.9 and an F1 score of 0.82, indicating that the present invention is more effective in detecting aerodynamic noise of wind turbine blades through adaptive detection and data enhancement driven strategies in the absence of abnormal samples. In addition, the present invention also has a smaller model size, making it more suitable for deployment in the online monitoring system of the health status of wind turbine blades.
[0049] The above-mentioned specific implementation can be partially adjusted in different ways by those skilled in the art without departing from the principle and purpose of the present invention. The protection scope of the present invention shall be based on the claims and shall not be limited by the above-mentioned specific implementation. Each implementation scheme within its scope shall be subject to the constraints of the present invention.
Claims
1. A method for monitoring the health status of wind turbine blades based on aerodynamic noise analysis, characterized in that: The aerodynamic noise signal of each blade is collected and the corresponding energy spectrum vector is generated, which is then combined with the energy spectrum vector data under historical normal working conditions to obtain an energy spectrum vector cluster. For each blade energy spectrum vector cluster, the DBSCAN clustering algorithm is used to perform online adaptive diagnosis based on unsupervised learning, identify abnormal noise points, and issue abnormal working condition warnings. The similarity refers to: the sum of the Euclidean distances of the energy spectrum vector of the current leaf relative to other energy spectrum vectors; The energy spectrum vector is obtained in the following way: a) performing analog-to-digital conversion on the collected aerodynamic noise signal of each blade and further performing low-pass digital filtering; b) performing adaptive signal variable mode decomposition on the blade aerodynamic noise signal obtained in step a; c) extracting the modal functions of the first several scales of the signal after decomposition in step b, calculating the energy of each modal function and the energy percentage representation of each modal function, and then obtaining an energy spectrum vector composed of the energy percentage representation.
2. The method for monitoring the health status of wind turbine blades according to claim 1, characterized in that: The collection mentioned above refers to: deploying sensor units on fan blades: for each fan, three sensor units are required to sample the aerodynamic noise signals when the blades are running. Since the aerodynamic noise generated when the fan blades are running is mainly distributed at the end of the blades, the sensors should be deployed at a suitable position near the end of the blades.
3. The method for monitoring the health status of wind turbine blades according to claim 1, characterized in that: The low-pass digital filtering, i.e. filtering out high-frequency noise from the collected data, specifically includes: 2.1 Construct a low-pass filter whose transfer function is: Among them: s is the differential operator, ω c is the filter cutoff angular frequency, ω under different sampling environments c different; 2.2 The low-pass filter transfer function is transformed through Tuspin to obtain a discretized differential equation, and the collected data is converted into digital form and then further low-pass digitally filtered.
4. The method for monitoring the health status of wind turbine blades according to claim 1, characterized in that: The adaptive signal variable mode decomposition specifically includes: 3.1 Find the maximum and minimum points of X(t), and use cubic spline interpolation to fit the envelope u(t) and v(t); 3.2 Subtract its own mean envelope from X(t), specifically: 3.3 Determine whether the number of extreme points and zero-crossing points in h(t) satisfies |ab|≤1 and whether the two envelopes of h(t) are symmetrical based on the horizontal coordinate. If so, take h(t) as the first mode function IMF1(t). Otherwise, let X(t) = h(t) and repeat step 2, where: a is the number of extreme points in h(t), and b is the number of zero-crossing points in h(t); 3.4 Subtract the first modal function IMF1(t) from X(t) to obtain the first-order residual signal r1(t), that is, r1(t) = X(t) - IMF1(t). Take r1(t) as the new signal to be processed and repeat steps 3.1 to 3.3 to obtain a series of modal functions IMF1(t), IMF2(t), ...IMF n (t) and the residual signals r1(t),r2(t),...r n (t), when the IMF n (t) The corresponding residual signal r n (t) is less than the threshold value, and the adaptive variable mode decomposition process ends; 3.5 Through adaptive variable mode decomposition, the original blade aerodynamic noise signal is represented as a series of mode functions IMF1(t), IMF2(t), ...IMF n (t) and the residual signals r1(t),r2(t),...r n The sum of (t) is 5. The method for monitoring the health status of wind turbine blades according to claim 1, characterized in that: The extraction refers to: extracting the modal functions IMF1(t) to IMF2(t) of the first N scales of the decomposed signal. N (t), extract the energy spectrum vector as a feature, and obtain the energy of each modal function through the discrete signal energy formula and Parseval theorem The energy percentage of each mode function is expressed as and the energy spectrum vector H p =[E p1 ,E p2 ,...,E pN ] T , where: T is the total data point length of mode function i.
6. The method for monitoring the health status of wind turbine blades according to claim 5 is characterized in that: The energy percentage representation, that is, the energy percentage of each modal function, represents the composition of modal components with different characteristics in the blade aerodynamic noise signal, and can effectively represent and distinguish blade audio signals with different aerodynamic noise characteristics.
7. The method for monitoring the health status of wind turbine blades according to claim 1, 5 or 6, characterized in that: The energy spectrum vector cluster is obtained by collecting the historical aerodynamic noise signals X of the three blades a, b, and c of each wind turbine under normal working conditions. a1 (t),X a2 (t),...,X am (t), X b1 (t),X b2 (t),...,X bm (t), X c1 (t),X c2 (t),...,X cm (t), perform signal processing from step 2 to step 4 on all historical aerodynamic noise signals to obtain a series of energy spectrum vectors H pa1 ,H pa2 ,...,H pam , H pb1 ,H pb2 ,...,H pbm , H pc1 ,H pc2 ,...,H pcm As data augmentation samples; Under online operation conditions, the aerodynamic noise signal x of the three blades of each wind turbine is a (t),x b (t),x c (t) Perform cross-correlation check, specifically: when any blade has a cross-correlation R with the other two blades xy If the value is smaller, an abnormal condition warning is issued for the blade, otherwise the next step is executed; specifically: Where: x, y are the aerodynamic noise signals of any two blades of the same wind turbine; The aerodynamic noise signal x of the blades of the same wind turbine a (t),x b (t),x c (t) Repeat low-pass digital filtering, adaptive signal modal decomposition and feature extraction to obtain the energy spectrum vector h of the aerodynamic noise signal of each blade a ,h b ,h c Then, it is combined with the energy spectrum vector data under historical normal working conditions to obtain the energy spectrum vector cluster: H a =[H pa1 ,H pa2 ,...,H pam ,h a ] T =[H' pa1 ,H' pa2 ,...,H' pam ,H' pa(m+1) ] T , H b =[H pb1 ,H pb2 ,...,H pbm ,h b ] T =[H' pb1 ,H' pb2 ,...,H' pbm ,H' pb(m+1) ] T , H c =[H pc1 ,H pc2 ,...,H pcm ,h c ] T =[H' pc1 ,H' pc2 ,...,H' pcm ,H' pc(m+1) ] T .
8. The method for monitoring the health status of wind turbine blades according to claim 1, 5 or 6, characterized in that: The online adaptive diagnosis based on unsupervised learning refers to: For each energy spectrum vector cluster of each blade, calculate the sum of the Euclidean distances of each energy spectrum vector relative to other energy spectrum vectors. For each blade, find the energy spectrum vector with the largest Euclidean distance relative to other energy spectrum vectors. Take the energy spectrum vector of blade a as the most recently sampled energy spectrum vector h. a When abnormal working condition is detected, the abnormal working condition diagnosis is performed based on the Euclidean distance criterion. Otherwise, the DBSCAN clustering algorithm is used to perform online adaptive diagnosis based on unsupervised learning for the energy spectrum vector cluster of each blade to identify abnormal noise points. The abnormal condition diagnosis based on the Euclidean distance criterion means: when the Eu of blade a a(m+1) ≥T a , when it is met, an abnormal working condition warning is issued to blade a. Otherwise, for the energy spectrum vector cluster of each blade, the DBSCAN clustering algorithm is applied to perform online adaptive diagnosis based on unsupervised learning to identify abnormal noise points.
9. The method for monitoring the health status of wind turbine blades according to claim 8, characterized in that: The DBSCAN clustering algorithm divides data points into core points, boundary points and abnormal noise points, where a core point means that there are at least MinPts other points in its ε-neighborhood, that is, N ε (p) = {q∈H a |d(p,q)≤ε}, where: d(p,q) is the distance between vector p and vector q. The condition for point p to be marked as a core point is |N ε (p)|≥MinPts; the boundary point refers to: when N ε If the number of points contained in (p) is less than MinPts, but p is within the ε-neighborhood of other core points, then p is a boundary point, otherwise it is an abnormal noise point.
10. A wind turbine blade health status monitoring system implementing the method described in any one of claims 1 to 9, characterized in that: include: A sampling unit, a feature extraction unit, a detection unit and an early warning unit, wherein: the sampling unit performs low-pass filtering and noise reduction preprocessing according to the aerodynamic noise audio information of the fan blades to obtain the aerodynamic noise audio signal after noise reduction; the feature extraction unit performs adaptive variable mode decomposition and energy spectrum calculation according to the aerodynamic noise audio signal after noise reduction to extract the energy spectrum vector corresponding to the aerodynamic noise audio signal; the detection unit performs similarity judgment, Euclidean distance calculation and DBSCAN cluster analysis according to the energy spectrum vectors corresponding to a series of aerodynamic noise audio signals obtained by sampling and feature extraction; the early warning unit issues an early warning on the health status of the corresponding blade according to the cluster analysis results and the similarity and Euclidean distance criteria.
