Wind turbine rotor abnormality detection method and system based on acoustic vibration signal identification

Abstract

The application discloses a wind turbine impeller abnormality detection method and system based on acoustic vibration signal recognition. The method first collects impeller response through a microphone / vibration sensor and generates a Mel spectrogram; then, Teager-Kaiser energy operators and autocorrelation analysis are used to identify the impeller rotation period and perform frequency spectrum segmentation without relying on the speed signal; the cross-period Mel spectrum dynamic deviation is calculated and normalized through the speed-related amplitude correction coefficient; the improved structural similarity algorithm is used to generate a periodic coherent energy map; and the harmonic resonance template is used for filtering enhancement, morphological deconstruction and parameterization of the abnormal area in the map, and geometric features are extracted; finally, the comprehensive abnormality score is calculated based on multi-dimensional features, and automatic early warning is realized. The application effectively overcomes the variable working condition interference and speed dependence problem, and can realize accurate and stable detection of early impeller abnormalities.

Description

Technical Field

[0001] This invention relates to the field of wind power equipment condition monitoring and fault diagnosis technology, specifically to a method and system for detecting abnormalities in wind turbine rotors based on acoustic and vibration signal recognition. Background Technology

[0002] As a key piece of equipment for clean energy, the long-term stable operation of wind turbines is crucial. The rotor (including blades and hub), as the core component for capturing wind energy, is subjected to complex alternating loads over long periods, making it susceptible to faults such as cracks, corrosion, leading-edge erosion, and lightning strike damage. These faults, in their early stages, alter the aerodynamic shape and structural dynamics of the blades, leading to subtle, patterned abnormal noise patterns during operation.

[0003] Traditional detection methods, such as vibration analysis and SCADA data analysis, are not sensitive to early, localized damage to blades. Vibration sensors are typically mounted on transmission chain components such as gearboxes and generators, making it difficult to effectively capture subtle fault characteristics originating from the blades. In recent years, acoustic-vibration-based monitoring technologies have attracted attention due to their advantages of being non-contact and providing rich information. However, the environmental noise in wind farms is complex, and blade noise is significantly affected by wind speed and rotational speed, causing abnormal acoustic-vibration characteristics to be easily masked, making stable and accurate early detection difficult. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and system for detecting wind turbine impeller anomalies based on acoustic vibration identification. This method extracts the blade rotation period using Mel spectrum analysis and autocorrelation methods, and calculates the dynamic deviation of the spectrum. It generates and enhances a periodic coherent energy map using an improved structural similarity algorithm and harmonic resonance template. Morphological analysis and feature extraction are performed on the abnormal regions in the map, ultimately achieving anomaly scoring and early warning based on multi-dimensional features. This method effectively overcomes interference from varying operating conditions and speed dependence, enabling accurate and stable detection of early impeller anomalies.

[0005] This invention is achieved through the following technical solution:

[0006] A method for detecting wind turbine rotor anomalies based on acoustic vibration signals includes the following steps:

[0007] (1) Microphones or microphone arrays and / or vibration sensors of corresponding protection levels are installed in the tower, nacelle, blade cavity and hub of the wind turbine to collect the sound and vibration signals of the wind turbine, and generate a two-dimensional Mel spectrum that reflects the hearing characteristics of the human ear through short-time Fourier transform and Mel frequency mapping.

[0008] (2) Without relying on the rotation speed signal, design an impeller rotation period identification method based on the time-frequency characteristic energy evolution of the impeller acoustic vibration signal to identify the impeller rotation period, and perform spectrum segmentation on the spectrum diagram according to the identified period;

[0009] (3) Calculate the Mel spectrum difference between different rotation period spectrum segments to obtain the cross-cycle Mel spectrum dynamic deviation (MSD), and use the amplitude correction coefficient related to the impeller speed to correct the aerodynamic noise amplitude of the impeller under different wind speeds to ensure that the abnormal characteristics are detectable under different working conditions.

[0010] (4) Design an improved structural similarity algorithm ISSIM, use ISSIM to measure the structural consistency of the dynamic deviation of the Mel spectrum, and obtain a periodic coherent energy map by introducing frequency band energy weight;

[0011] (5) The image domain of the periodic coherence energy map is filtered and enhanced using a harmonic resonance template. Based on the abnormal morphological deconstruction and parameterization method of the Mel spectrum deviation map, the frequency structure of the abnormal signal of the wind turbine impeller is accurately described. The periodic coherence map is used as input to perform multi-dimensional quantitative analysis on the morphological features of the abnormal region, thereby obtaining periodic feature parameters such as intensity, quantity or distribution for fault identification and pattern classification.

[0012] (6) Based on the cross-cycle Mel spectrum dynamic deviation obtained in step (2) and the periodic characteristic parameters obtained in step (5), the acoustic and vibration state of the wind turbine impeller is continuously monitored and anomalies are identified. The system calculates the energy distribution, structural similarity, etc. in each rotation cycle in real time, compares them with the baseline model during healthy operation, and calculates a comprehensive anomaly score based on these parameters. With the preset alarm threshold In comparison, It issues alarm signals in a timely manner. This mechanism enables early detection and automatic warning of abnormal impeller acoustic vibration, providing reliable support for wind turbine operating status assessment and fault prevention.

[0013] Furthermore, the blade rotation period identification method based on the time-frequency characteristic energy evolution of impeller noise in step (2) is specifically as follows:

[0014] The Mel spectrum is denoted as , indicating the first Frame, First The logarithmic energy of each Mel frequency band, through the... In the selected frequency band set Perform frequency domain integration to calculate the one-dimensional energy envelope. After processing with the Teager–Kaiser energy operator, the normalized autocorrelation function is expressed as follows:

[0015] ;

[0016] Determine the rotation period corresponding to the main peak Based on the identified period The continuous Mel spectrum is divided into several independent segments along the time axis. Each segment contains complete information of one blade rotation cycle, thereby realizing the periodic synchronization of subsequent feature extraction and fault identification.

[0017] in, For Mel frequency band index; To select the set of frequency bands for integration, the entire band or several bands near the blade pass frequency BPF are selected to improve the stability of the calculation results;

[0018] The one-dimensional energy envelope obtained by integrating the Mel spectrum in the frequency domain;

[0019] The Teager–Kaiser Energy Operator (TKEO) enhances the envelope, highlighting the "over-leaf" instantaneous modulation; it is defined as:

[0020] ;

[0021] For the estimated rotation period, in step size Segment the Mel spectrum.

[0022] Step (3) is specifically implemented through the following process:

[0023] After dividing the continuous acoustic vibration signal according to the rotation period, the first... The Mel spectrum of each period is represented as , Index the time frames within this period, using the reference spectrum in a healthy operating state. Based on this, define the periodic spectral difference. for:

[0024] ;

[0025] in, The total number of time frames within a period. For frequency band weighting coefficients, This is the norm exponent parameter (usually 1 or 2). The frequency band weighting coefficient... The frequency band energy distribution near the blade passing frequency (BPF) and its harmonics is determined, satisfying...

[0026] ;

[0027] in The harmonic order is... For the first The center frequency of each Mel band, It is the frequency of the blades passing through. This is a bandwidth adjustment parameter used to control the width of the harmonic resonant template around the harmonic frequency;

[0028] The amplitude correction coefficient in step (3) is defined as:

[0029] ;

[0030] in, This represents the angular velocity of the blade. Indicates the rated angular velocity. C It is an offset parameter that represents the center point of the speed ratio where the response changes most drastically (typically close to 1). K (m,n) is a scaling factor used to normalize the magnitude. This is an empirical power matrix, whose element values ​​are set according to the main sound source characteristics of its corresponding time-frequency unit (m,n):

[0031] ;

[0032] Frequency dimension (m Low-frequency noise related to blade thickness tends to exhibit dipole characteristics, while high-frequency noise related to tip vortices tends to exhibit quadrupole characteristics. Therefore, the matrix can be assigned a value of 2 in the corresponding low-frequency region and a value of 4-5 in the high-frequency region; time dimension ( n ): During a rotation cycle, when the blades pass through the tower or when the pitch angle changes, specific noises may be generated;

[0033] The formula for calculating Mel spectral dynamic deviation (MSD) is as follows:

[0034] ,,in This represents the difference in the periodic spectrum.

[0035] Furthermore, in step (4), the ISSIM is used to measure the consistency of the dynamic deviation structure of the Mel spectrum, and a periodic coherent energy map is obtained by introducing frequency band energy weights, specifically including:

[0036] Introducing energy weights The improved structural similarity algorithm ISSIM is used to calculate the structural similarity as follows:

[0037] ;in, and The Mel spectrum deviation diagrams are for two adjacent rotation cycles, respectively. , and These represent the mean, standard deviation, and covariance, respectively. It is a constant used to prevent the denominator from being zero; Here, is the energy weighting coefficient, where For the first The total spectral energy of the period, It is the average total spectral energy, used to enhance the contribution of high-energy frequency bands; the ISSIM is an improved algorithm optimized to address the problem that the traditional SSIM algorithm is sensitive to amplitude drift and random noise and has difficulty accurately reflecting the periodic structural changes of the blade, and is used to calculate the structural consistency of the cross-period Mel spectrum deviation map.

[0038] Furthermore, in step (5), the periodic coherent energy map is filtered and enhanced using a harmonic resonance template in the image domain. This is achieved through the following process:

[0039] Periodic coherent energy maps are generated using the structural similarity results output by the ISSIM algorithm. Harmonic resonance morphological filtering is then applied in the image domain to enhance the contrast of anomalous regions and suppress non-periodic noise. The calculation steps are as follows:

[0040] This method first constructs a harmonic resonance template. H(m) This is used to amplify true periodic anomalous signals on a periodic coherent energy map while filtering out random noise. The expression is as follows:

[0041] ;

[0042] in, m It is a Mel frequency band index. It is the first m The center frequency of each Mel frequency band h It is the harmonic order, from 1 to H .

[0043] Next, harmonic resonance morphological filtering is performed, assuming the input "periodic coherent energy map" is... ,in m It is a frequency band index. n This is the time frame index, the filtered output:

[0044] ;

[0045] It is the original image The result after applying local minimum filtering. The values ​​of the harmonic resonance template to be created are limited to [0, 1].

[0046] Furthermore, the abnormal morphological deconstruction and parameterization method in step (5) is specifically as follows:

[0047] This includes extracting the topological boundaries of highlighted connected components in a periodic coherent energy map based on the Canny edge detection algorithm, and calculating the center frequency of each boundary profile. ,bandwidth curvature and skewness With equal geometric parameters, the following multidimensional feature vectors are formed:

[0048]

[0049] Among them, center frequency Characterizes the principal energy location of the anomaly in the frequency domain, and the frequency corresponding to the centroid of the frequency band index of all pixels within the contour; bandwidth. Frequency range reflecting energy distribution; curvature Describe the degree of curvature of the abnormal profile; skewness. This indicates the asymmetry of its frequency band energy distribution.

[0050] Furthermore, to correct the bandwidth deviation caused by edge detection, an error compensation coefficient is introduced. The revised bandwidth calculation formula is as follows: ; in This can be obtained through statistical bias analysis of healthy samples, and used to compensate for bandwidth bias caused by changes in signal-to-noise ratio or boundary blurring. The corrected feature vector. It can accurately characterize the frequency range and morphological features of anomalies, providing a quantifiable basis for the subsequent identification, clustering and classification of abnormal acoustic patterns in blades.

[0051] Furthermore, step (6) is specifically implemented through the following process:

[0052] Real-time analysis of dynamic deviations, structural similarities, and morphological parameters of spectra across multiple consecutive periods is performed to calculate a comprehensive anomaly score. The stability and anomalous trends of the acoustic-vibration signal are assessed by comparing it with a healthy baseline model. Specifically, the comprehensive anomaly score is calculated by integrating the statistics of the cross-cycle Mel-frequency dynamic deviation (MSD), the degree of decrease in ISSIM, and the evolution of the eigenvector F, using a pre-trained support vector machine (SVM) model or a weighted summation method. ;

[0053] When the comprehensive abnormal score Exceeding the preset threshold At that time, that is It will automatically generate alarm signals to realize online early warning and operation status prompts for blade anomalies, and maintain the amplitude consistency of abnormal sound and vibration characteristics and the stability of blade anomaly identification and diagnosis under different wind speeds and operation conditions.

[0054] This invention also discloses a wind turbine rotor anomaly detection system based on acoustic vibration recognition, which includes the following units:

[0055] Acquisition and mapping unit: Microphones or microphone arrays and / or vibration sensors with corresponding protection levels are installed in the wind turbine tower, nacelle, blade cavity and hub to collect the sound and vibration signals of the wind turbine, and generate a two-dimensional Mel spectrum that reflects the characteristics of human hearing through short-time Fourier transform and Mel frequency mapping.

[0056] Spectrum Segmentation Unit: Without relying on the rotational speed signal, a method for identifying the impeller rotation period based on the energy evolution of the time-frequency characteristics of impeller acoustic vibration is designed to identify the impeller rotation period, and the spectrum diagram is segmented according to the identified period;

[0057] Amplitude correction unit: Calculates the Mel spectrum difference between different rotation cycle spectrum segments to obtain the cross-cycle Mel spectrum dynamic deviation, and uses an amplitude correction coefficient related to impeller speed to correct the aerodynamic noise amplitude of the impeller under different wind speeds to ensure that abnormal characteristics are detectable under different operating conditions;

[0058] Obtaining periodic coherent energy map units: An improved structural similarity algorithm ISSIM is designed to measure the structural consistency of the Mel spectrum dynamic deviation. By introducing frequency band energy weights, a periodic coherent energy map is obtained.

[0059] Multi-dimensional analysis unit: The harmonic resonance template is used to filter and enhance the image domain of the periodic coherent energy map, and the frequency structure of the abnormal signal of the wind turbine blade is accurately described based on the abnormal morphological deconstruction and parameterization method of Mel spectrum deviation map. With the periodic coherent energy map as input, the morphological features of the abnormal region are analyzed in multiple dimensions to obtain periodic feature parameters for fault identification and pattern classification.

[0060] Automatic alarm unit: Based on the cross-cycle Mel spectrum dynamic deviation obtained by the spectrum segmentation unit and the periodic characteristic parameters obtained by the multi-dimensional analysis unit, the sound and vibration status of the wind turbine impeller is continuously monitored and anomalies are identified. Finally, a comprehensive anomaly score is calculated based on the multi-dimensional features. When the comprehensive anomaly score is greater than the preset alarm threshold, an alarm signal is issued to achieve automatic early warning.

[0061] The technical solutions provided by the embodiments of this application may include the following beneficial effects:

[0062] Independent of speed signal: Through innovative time-frequency energy evolution analysis, the blade rotation cycle can be accurately identified without speed sensor, reducing system cost and deployment complexity.

[0063] Strong environmental robustness: Through amplitude correction coefficient and harmonic resonance morphological filtering, the interference of wind speed changes and random noise is effectively suppressed, ensuring the stability and sensitivity of anomaly detection under different operating conditions.

[0064] Deep feature extraction: Combining signal processing and image analysis techniques, it refines the deconstruction of anomalies from multiple dimensions such as dynamic deviation, structural consistency, and geometric morphology, resulting in stronger feature expression capabilities and facilitating the identification of early-stage weak faults and complex fault modes.

[0065] Automation and intelligence: The entire process, from signal acquisition to alarm output, is fully automated. Through a comprehensive anomaly scoring mechanism, manual intervention is reduced, and the level of intelligence in wind turbine condition monitoring and operation and maintenance efficiency are improved. Attached Figure Description

[0066] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0067] Figure 1 This is a flowchart illustrating the identification and segmentation of blade rotation period without speed signal in a wind turbine impeller anomaly detection method based on acoustic vibration signals, according to an exemplary embodiment.

[0068] Figure 2 This is a flowchart illustrating the abnormal feature extraction and state assessment process in a wind turbine impeller anomaly detection method based on acoustic and vibration signals, according to an exemplary embodiment. Detailed Implementation

[0069] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention as detailed in the appended claims.

[0070] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application.

[0071] This embodiment provides a method for detecting wind turbine impeller anomalies based on acoustic and vibration signal recognition, such as... Figure 1 As shown, the specific steps include: Step 1: Acoustic signal acquisition and auditory spectrum generation. High-precision, protected microphones or microphone arrays and / or vibration sensors are installed in the wind turbine tower, nacelle, blade cavity, and hub to acquire the acoustic and vibration signals of the wind turbine. Through short-time Fourier transform and Mel frequency mapping, a two-dimensional Mel spectrum diagram reflecting the characteristics of human hearing is generated. The specific implementation process is as follows:

[0072] A high-precision microphone meeting the required protection level is installed on the outer side of the bottom of the wind turbine tower to collect rotor rotation noise. The sampling frequency is set to 44.1kHz. The collected raw sound pressure signal is processed in frames (1024 points per frame, 50% overlap), and a short-time Fourier transform is performed on each frame to obtain its power spectrum. Subsequently, the power spectrum is passed through a Mel filter bank containing 128 filters, and the logarithmic energy of each filter output is calculated, finally generating a 128-dimensional Mel frequency cepstral coefficient (MFCC) spectrum, namely the Mel spectrum S(m, t) described in this invention, where m=1,2,...,128 is the frequency band index, and t is the time frame index.

[0073] Step 2: Rotation Period Identification and Signal Segmentation without Rotation Speed ​​Signal. Without relying on the rotation speed signal, a method for identifying the impeller rotation period based on the energy evolution of the impeller's acoustic vibration time-frequency characteristics is designed. The method then segments the spectrum based on the identified period. Figure 2 As shown, the specific implementation process is as follows:

[0074] Select Mel spectrum (indicating the first) Frame, First The one-dimensional energy envelope is obtained by integrating the logarithmic energy of each Mel frequency band into the mid-to-low frequency bands (e.g., the first 64 frequency bands) that are related to the blade's passing frequency. ;

[0075] Using the Teager-Kaiser energy operator (TKEO) Enhancement: This operator can effectively amplify the instantaneous energy impact generated when the blades sweep across the tower.

[0076] calculate Normalized autocorrelation function Based on the fan model, set the cycle search window to [0.5s, 5.0s] (corresponding to a speed range of 12RPM to 120RPM); search within this window. The maximum value of τ corresponds to the estimated rotation period. ; where the normalized autocorrelation function The expression is as follows:

[0077] ;

[0078] Based on the estimated rotation period , to continuously display the Mel spectrum Divide into K periodic segments , k=1,2,... ,K.

[0079] Step 3: Cross-cycle spectral deviation calculation and amplitude correction. The Mel spectrum difference between different rotational cycle spectral segments is calculated to obtain the cross-cycle Mel spectrum dynamic deviation (MSD). An amplitude correction coefficient related to impeller speed is then used to correct the impeller aerodynamic noise amplitude under different wind speeds, ensuring that abnormal characteristics are detectable under different operating conditions. The specific implementation process is as follows:

[0080] The average spectrum was established based on data collected during the initial stage of the unit's healthy operation. Using the reference spectrum as a baseline, calculate the difference between the k-th period segment and the reference spectrum: ;

[0081] in, The total number of time frames within a period. For frequency band weighting coefficients, The norm exponent parameter is typically 1 or 2; the frequency band weighting coefficient... Determined based on the frequency band energy distribution near the blade's passing frequency BPF and its harmonics, satisfying:

[0082] ;

[0083] in The harmonic order is... For the first The center frequency of each Mel band, It is the frequency of the blades passing through. This is a bandwidth adjustment parameter used to control the width of the harmonic resonant template around the harmonic frequency.

[0084] Amplitude correction factor Defined as: ;in, The estimated angular velocity for the current period. For the rated angular velocity, C is taken as 0.9-1.1, K is the normalization factor, and β is the parameter matrix set according to the characteristics of the frequency band sound source (takes 2 for the low-frequency dipole region and 4 for the high-frequency quadrupole region).

[0085] The corrected dynamic bias of the Mel spectrum is: .

[0086] Step 4: Improve structural similarity measurement and periodic coherence map generation. An improved structural similarity algorithm, ISSIM, is designed to measure the structural consistency of the Mel spectrum dynamic deviation. By introducing frequency band energy weights, a periodic coherence energy map is generated using the improved structural similarity algorithm. The specific implementation process is as follows:

[0087] by and As input, calculate its improved structural similarity: ; in, and The Mel spectrum deviation diagrams are for two adjacent rotation cycles, respectively. , and These represent the mean, standard deviation, and covariance, respectively. It is a constant used to prevent the denominator from being zero; Here, the energy weighting coefficient is... For the first The total spectral energy of the period, It is the average total spectral energy, used to enhance the contribution of high-energy frequency bands; the ISSIM is an improved algorithm optimized to address the problem that the traditional SSIM algorithm is sensitive to amplitude drift and random noise and has difficulty accurately reflecting the periodic structural changes of the blade, and is used to calculate the structural consistency of the cross-period Mel spectrum deviation map.

[0088] Arrange the continuously calculated ISSIM values ​​in chronological order to form an initial periodic coherent energy map C(m, n);

[0089] Constructing a harmonic resonance template Where m is the Mel frequency band index, It is the center frequency of the m-th Mel frequency band. is the highest harmonic order, and h is the harmonic order, from 1 to H;

[0090] Perform harmonic resonance morphological filtering: ,in for The results of local minimum filtering are shown. After filtering, noise in the non-harmonic region is effectively suppressed, while the abnormal region related to the blade passage frequency harmonics is highlighted.

[0091] Step 5: Anomaly Morphological Deconstruction and Parameterization. Harmonic resonance templates are used for filtering and enhancement in the image domain of the periodicity coherence map. Based on the Mel spectrum deviation map-based anomaly morphological deconstruction and parameterization method, the frequency structure of the wind turbine impeller anomaly signal is accurately described. Using the periodicity coherence map as input, multi-dimensional quantitative analysis is performed on the morphological features of the anomaly regions to obtain periodic feature parameters such as intensity, quantity, or distribution for fault identification and pattern classification. The specific implementation process is as follows:

[0092] Periodic coherent energy maps are generated using the structural similarity results output by the ISSIM algorithm. Harmonic resonance morphological filtering is then applied in the image domain to enhance the contrast of anomalous regions and suppress non-periodic noise. The calculation steps are as follows:

[0093] This method first constructs a harmonic resonance template H(m) to amplify the true periodic anomalous signal on the periodic coherent energy map, while filtering out random noise. The expression is as follows:

[0094] ;

[0095] Where m is the Mel frequency band index, It is the center frequency of the m-th Mel frequency band, and h is the harmonic order, from 1 to H.

[0096] Next, harmonic resonance morphological filtering is performed, assuming the input "periodic coherent energy map" is... Where m is the frequency band index and n is the time frame index, the filtered output is:

[0097] ;

[0098] It is the original image The result after applying local minimum filtering. The values ​​of the harmonic resonance template to be created are limited to [0, 1].

[0099] The filtered coherent image is binarized, and the Canny operator is used to extract the contours of anomalous regions. Then, using the periodic coherent energy map as input, the topological boundaries of the highlighted connected components in the image are extracted based on the Canny edge detection algorithm. Geometric parameters such as center frequency, bandwidth, curvature, and skewness are calculated for each boundary contour, forming the following multidimensional feature vector:

[0100] ;

[0101] For each connected component contour, the following feature parameters are calculated:

[0102] Center frequency : Characterizes the principal energy location of the anomaly in the frequency domain, and the frequency corresponding to the centroid of the frequency band index of all pixels within the contour;

[0103] bandwidth The range spanned by the profile on the frequency axis. To compensate for detection errors, a compensation coefficient γ_c is introduced. The corrected bandwidth calculation formula is:

[0104] ;in This can be obtained through statistical bias analysis of healthy samples, and used to compensate for bandwidth bias caused by changes in signal-to-noise ratio or boundary blurring. The corrected feature vector. It can accurately characterize the frequency range and morphological features of anomalies, providing a quantifiable basis for the subsequent identification, clustering and classification of abnormal acoustic vibration modes of blades;

[0105] curvature : The average curvature that describes the degree of bending of the profile;

[0106] Skewness The asymmetry of the profile energy distribution along the frequency axis.

[0107] Step 6: Comprehensive Status Assessment and Automatic Early Warning. Based on the cross-cycle Mel spectrum dynamic deviation from step (2) and the periodic characteristic parameters obtained from step (5), the acoustic and vibration status of the wind turbine blades is continuously monitored and anomalies are identified. The system calculates the energy distribution, structural similarity, etc., within each rotation cycle in real time, compares them with the baseline model during healthy operation, and calculates a comprehensive anomaly score based on these parameters. With the preset alarm threshold In comparison, An alarm signal is issued in a timely manner. This mechanism enables early detection and automatic warning of abnormal blade acoustic vibration, providing reliable support for wind turbine operating status assessment and fault prevention. The specific implementation process is as follows:

[0108] By combining statistics of the cross-period Mel spectral dynamic deviation (MSD), such as mean and variance, the degree of decrease in ISSIM, and the evolution of the eigenvector F, a comprehensive anomaly score between 0 and 100 is calculated using a pre-trained support vector machine (SVM) model or a weighted summation method. .

[0109] when Exceeding the preset threshold for multiple consecutive cycles (e.g., 5 cycles) When the parameter is typically set to 75, the system triggers an alarm and sends the alarm information and a detailed diagnostic report to the maintenance personnel via the network.

[0110] The method described in this invention, through the organic combination of the above steps, achieves early, stable, automated, and intelligent detection of abnormal acoustic vibration of wind turbine rotors.

[0111] This embodiment also discloses a wind turbine rotor anomaly detection system based on acoustic vibration recognition, which includes the following units:

[0112] Acquisition and mapping unit: Microphones or microphone arrays and / or vibration sensors with corresponding protection levels are installed in the wind turbine tower, nacelle, blade cavity and hub to collect the sound and vibration signals of the wind turbine, and generate a two-dimensional Mel spectrum that reflects the characteristics of human hearing through short-time Fourier transform and Mel frequency mapping.

[0113] Spectrum Segmentation Unit: Without relying on the rotational speed signal, a method for identifying the impeller rotation period based on the energy evolution of the time-frequency characteristics of blade noise is designed to identify the impeller rotation period, and the spectrum diagram is segmented according to the identified period;

[0114] Amplitude correction unit: Calculates the Mel spectrum difference between different rotation cycle spectrum segments to obtain the cross-cycle Mel spectrum dynamic deviation, and uses an amplitude correction coefficient related to impeller speed to correct the aerodynamic noise amplitude of the impeller under different wind speeds to ensure that abnormal characteristics are detectable under different operating conditions;

[0115] Obtaining periodic coherent energy map units: Based on the similarity algorithm SSIM, an improved structural similarity algorithm ISSIM is designed. ISSIM is used to measure the structural consistency of the dynamic deviation of the Mel spectrum, and by introducing frequency band energy weights, the periodic coherent energy map is obtained.

[0116] Multi-dimensional analysis unit: The harmonic resonance template is used to filter and enhance the image domain of the periodic coherent energy map, and the frequency structure of the abnormal signal of the wind turbine blade is accurately described based on the abnormal morphological deconstruction and parameterization method of Mel spectrum deviation map. With the periodic coherent energy map as input, the morphological features of the abnormal region are analyzed in multiple dimensions to obtain periodic feature parameters for fault identification and pattern classification.

[0117] Automatic alarm unit: Based on the cross-cycle Mel spectrum dynamic deviation obtained by the spectrum segmentation unit and the periodic characteristic parameters obtained by the multi-dimensional analysis unit, the sound and vibration status of the wind turbine blades is continuously monitored and anomalies are identified. Finally, a comprehensive anomaly score is calculated based on the multi-dimensional features. When the comprehensive anomaly score is greater than the preset alarm threshold, an alarm signal is issued to achieve automatic early warning.

[0118] In summary, this invention extracts the blade rotation period and calculates the spectral dynamic deviation using Mel spectrum analysis and autocorrelation methods; it generates and enhances the periodic coherent energy map using an improved structural similarity algorithm and harmonic resonance template; and it achieves anomaly scoring and early warning based on multi-dimensional features by performing morphological analysis and feature extraction on anomalous regions in the map. Compared with traditional methods that rely on only a single or dual feature dimension (such as only center frequency and bandwidth) for anomaly characterization, the four-dimensional feature vector proposed in this invention achieves a more comprehensive and refined characterization of anomalous blade acoustic vibration modes by integrating four complementary dimensions: frequency domain position, bandwidth robustness, morphological curvature, and distribution symmetry. This multi-dimensional feature system not only significantly improves the discriminability and identifiability of anomalous modes but also effectively addresses feature drift and interference under complex operating conditions, providing quantifiable and highly reliable feature basis for subsequent accurate identification, clustering, and classification of blade anomalies, thus demonstrating significant advantages in diagnostic accuracy and robustness.

[0119] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method for detecting wind turbine impeller anomalies based on acoustic and vibration signal recognition, characterized in that, The following methods and steps are included: (1) Microphones or microphone arrays and / or vibration sensors of corresponding protection levels are installed in the tower, nacelle, blade cavity and hub of the wind turbine to collect the sound and vibration signals of the wind turbine, and generate a two-dimensional Mel spectrum that reflects the hearing characteristics of the human ear through short-time Fourier transform and Mel frequency mapping. (2) Design a method for identifying the impeller rotation period based on the energy evolution of the time-frequency characteristics of the impeller acoustic vibration signal, and segment the spectrum according to the identified period; (3) Calculate the Mel spectrum difference between different rotation cycle spectrum segments to obtain the cross-cycle Mel spectrum dynamic deviation, and use the amplitude correction coefficient related to the impeller speed to correct the aerodynamic noise amplitude of the impeller under different wind speeds to ensure that the abnormal characteristics are detectable under different working conditions. (4) Design an improved structural similarity algorithm ISSIM, use ISSIM to measure the structural consistency of the dynamic deviation of the Mel spectrum, and obtain the periodic coherent energy map by introducing frequency band energy weight; The process of obtaining a periodic coherent energy map by introducing frequency band energy weights specifically includes: Introducing energy weights The improved structural similarity algorithm ISSIM is used to calculate the structural similarity as follows: ; in, and The Mel spectrum deviation diagrams are for two adjacent rotation cycles, respectively. , and These represent the mean, standard deviation, and covariance, respectively. It is a constant used to prevent the denominator from being zero; Here, is the energy weighting coefficient, where For the first The total spectral energy of the period, It is the average total spectral energy, used to enhance the contribution of high-energy frequency bands; the ISSIM is an improved algorithm optimized from the traditional SSIM algorithm, used to calculate the structural consistency of the cross-cycle Mel spectral deviation map; (5) The image domain of the periodic coherent energy map is filtered and enhanced by the harmonic resonance template. Based on the abnormal morphological deconstruction and parameterization method of the Mel spectrum deviation map, the frequency structure of the abnormal signal of the wind turbine impeller is accurately described. The periodic coherent energy map is used as input to perform multi-dimensional quantitative analysis on the morphological features of the abnormal region, so as to obtain periodic feature parameters for fault identification and pattern classification. (6) Based on the cross-cycle Mel spectrum dynamic deviation obtained in step (2) and the periodic characteristic parameters obtained in step (5), the acoustic vibration state of the wind turbine impeller is continuously monitored and anomaly is identified. Finally, a comprehensive anomaly score is calculated based on multi-dimensional characteristics. When the comprehensive anomaly score is greater than the preset alarm threshold, an alarm signal is issued to realize automatic early warning.

2. The method for detecting wind turbine impeller anomalies based on acoustic and vibration signal recognition according to claim 1, characterized in that, The blade rotation period identification method based on the energy evolution of impeller acoustic vibration time-frequency characteristics in step (2) is as follows: The Mel spectrum is denoted as , indicating the first Frame, First The logarithmic energy of each Mel frequency band, through the... In the selected frequency band set Perform frequency domain integration to calculate the one-dimensional energy envelope. After processing with the Teager–Kaiser energy operator, the normalized autocorrelation function is expressed as follows: ； Determine the rotation period corresponding to the main peak Based on the identified period The continuous Mel spectrum is divided into several independent segments along the time axis, each segment containing complete information of one blade rotation cycle, thereby achieving periodic synchronization of subsequent feature extraction and fault identification; among them, For Mel frequency band index; To select the set of frequency bands for integration; This is a one-dimensional energy envelope obtained by integrating the Mel spectrum in the frequency domain. The Teager–Kaiser energy operator TKEO enhances the envelope, highlighting the "leaf-crossing" instantaneous modulation; defined as: ; For the feature time series used in periodic detection, we take... ; for The time mean; The physically feasible periodic search window is calculated from the known blade rotation speed range. For the estimated rotation period, in step size Segment the Mel spectrum.

3. The method for detecting wind turbine impeller anomalies based on acoustic and vibration signal recognition according to claim 1, characterized in that, Step (3) is specifically implemented through the following process: After dividing the continuous acoustic vibration signal according to the rotation period, the first... The Mel spectrum of each period is represented as , Index the time frames within this period, using the reference spectrum in a healthy operating state. Based on this, define the periodic spectral difference. for: ； in, The total number of time frames within a period. For frequency band weighting coefficients, The norm exponent parameter is 1 or 2; the frequency band weighting coefficient Determined based on the frequency band energy distribution near the blade's passing frequency BPF and its harmonics, satisfying: ； in Where h is the highest harmonic order, and h is the harmonic order. For the first The center frequency of each Mel band, It is the frequency of the blades passing through. This is a bandwidth adjustment parameter used to control the width of the harmonic resonant template around the harmonic frequency.

4. The method for detecting wind turbine impeller anomalies based on acoustic and vibration signal recognition according to claim 1, characterized in that, The amplitude correction coefficient in step (3) is defined as: ； in, This represents the angular velocity of the blade. This represents the rated angular velocity, C is an offset parameter representing the center point of the speed ratio where the response changes most drastically, and K(m,n) is a scaling factor used to normalize the amplitude. This is an empirical power matrix, whose element values ​​are set according to the main sound source characteristics of its corresponding time-frequency unit (m,n):

5. For the frequency dimension (m): The low-frequency noise generated by blade thickness disturbance has acoustic characteristics that conform to the characteristics of a dipole sound source; while the high-frequency noise generated by tip vortices has acoustic characteristics that conform to the characteristics of a quadrupole sound source; therefore, the matrix is ​​assigned a value of 2 in the corresponding low-frequency band region and a value of 4-5 in the high-frequency band region; For the time dimension (n): Within one rotation cycle, specific noise will be excited when the blade passes through the tower or when the pitch angle changes; The formula for calculating the dynamic deviation of the Mel spectrum is: ,in This represents the difference in the periodic spectrum.

6. The wind turbine impeller anomaly detection method based on acoustic and vibration signal recognition according to claim 1, characterized in that, In step (5), the periodic coherent energy map is filtered and enhanced using a harmonic resonance template in the image domain. This is achieved through the following process: A periodic coherent energy map is generated using the structural similarity results output by the ISSIM algorithm. Harmonic resonance morphological filtering is then applied in the image domain to enhance the contrast of anomalous regions and suppress non-periodic noise. Specifically, a harmonic resonance template H(m) is first constructed to amplify the true periodic anomalous signal on the periodic coherent energy map while filtering out random noise. The expression is as follows: ； Where m is the Mel frequency band index, is the center frequency of the m-th Mel frequency band, and h is the harmonic order, from 1 to H; Next, harmonic resonance morphological filtering is performed, assuming the input "periodic coherent energy map" is... Where m is the frequency band index and n is the time frame index, the filtered output is: ； It is the original image The result after applying local minimum filtering. The values ​​of the harmonic resonance template to be created are limited to [0, 1].

7. The method for detecting wind turbine impeller anomalies based on acoustic and vibration signal recognition according to claim 1, characterized in that, The abnormal morphological deconstruction and parameterization method in step (5) specifically includes: extracting the topological boundaries of the highlighted connected domains in the periodic coherent energy map based on the Canny edge detection algorithm, and calculating the center frequency of each boundary contour. ,bandwidth curvature and skewness Geometric parameters, forming the following multidimensional feature vector: ； Among them, center frequency Characterizes the principal energy location of the anomaly in the frequency domain, and the frequency corresponding to the centroid of the frequency band index of all pixels within the contour; bandwidth. Frequency range reflecting energy distribution; curvature Describe the degree of curvature of the abnormal profile; skewness. This indicates the asymmetry of its frequency band energy distribution.

8. The method for detecting wind turbine impeller anomalies based on acoustic and vibration signal recognition according to claim 6, characterized in that, To correct the bandwidth deviation caused by edge detection, an error compensation coefficient is introduced. The revised bandwidth calculation formula is as follows: ; in This can be obtained through statistical bias of healthy samples, and is used to compensate for bandwidth bias caused by changes in signal-to-noise ratio or boundary ambiguity. The corrected feature vector is then used. It can accurately characterize the frequency range and morphological features of anomalies, providing a quantifiable basis for the subsequent identification, clustering and classification of abnormal acoustic vibration modes of blades.

9. The method for detecting wind turbine impeller anomalies based on acoustic and vibration signal recognition according to claim 1. Its features are, Step (6) is specifically implemented through the following process: Real-time analysis of spectral dynamic deviation, structural similarity, and morphological parameters across multiple consecutive cycles is performed to continuously monitor and identify anomalies in the acoustic and vibration state of the wind turbine impeller. Energy distribution and structural similarity within each rotation cycle are calculated in real time and compared with a healthy baseline model to assess signal stability and anomaly trends. Specifically, a comprehensive anomaly score is calculated by integrating the statistics of the cross-cycle Mel spectral dynamic deviation (MSD), the degree of decrease in ISSIM, and the evolution of the eigenvector F, using a pre-trained support vector machine (SVM) model or a weighted summation method. ; When the comprehensive abnormal score Exceeding the preset threshold At that time, that is It will automatically generate alarm signals to realize online early warning and operation status prompts for blade anomalies, and maintain the amplitude consistency of abnormal sound and vibration characteristics and the stability of blade anomaly identification and diagnosis under different wind speeds and operation conditions.

10. A system for detecting wind turbine rotor anomalies based on acoustic and vibration signal recognition as described in any one of claims 1-8, characterized in that, The system includes the following units: Acquisition and mapping unit: Microphones or microphone arrays and / or vibration sensors with corresponding protection levels are installed in the wind turbine tower, nacelle, blade cavity and hub to collect the sound and vibration signals of the wind turbine, and generate a two-dimensional Mel spectrum that reflects the characteristics of human hearing through short-time Fourier transform and Mel frequency mapping. Spectrum Segmentation Unit: Without relying on the rotational speed signal, a method for identifying the impeller rotation period based on the energy evolution of the time-frequency characteristics of impeller acoustic vibration is designed to identify the impeller rotation period, and the spectrum diagram is segmented according to the identified period; Amplitude correction unit: Calculates the Mel spectrum difference between different rotation cycle spectrum segments to obtain the cross-cycle Mel spectrum dynamic deviation, and uses an amplitude correction coefficient related to impeller speed to correct the aerodynamic noise amplitude of the impeller under different wind speeds to ensure that abnormal characteristics are detectable under different operating conditions; Obtaining periodic coherent energy map units: An improved structural similarity algorithm ISSIM is designed to measure the structural consistency of the Mel spectrum dynamic deviation. By introducing frequency band energy weights, a periodic coherent energy map is obtained. Multi-dimensional analysis unit: The image domain of the periodic coherent energy map is filtered and enhanced using a harmonic resonance template. Based on the abnormal morphological deconstruction and parameterization method of Mel spectrum deviation map, the frequency structure of the abnormal signal of the wind turbine blade is accurately described. With the periodic coherent energy map as input, the morphological features of the abnormal region are analyzed in multiple dimensions to obtain periodic feature parameters for fault identification and pattern classification. Automatic alarm unit: Based on the cross-cycle Mel spectrum dynamic deviation obtained by the spectrum segmentation unit and the periodic characteristic parameters obtained by the multi-dimensional analysis unit, the sound and vibration status of the wind turbine impeller is continuously monitored and anomalies are identified. Finally, a comprehensive anomaly score is calculated based on the multi-dimensional features. When the comprehensive anomaly score is greater than the preset alarm threshold, an alarm signal is issued to achieve automatic early warning.

Images