A wind turbine blade acoustic dual-path anomaly detection method and system

Abstract

The present application relates to a kind of wind turbine blade acoustic dual-path anomaly detection method and system, by collecting multiple sound signals and combining SCADA system screening effective data.Establish noise model, adopt dynamic spectrum subtraction for noise suppression, then time alignment is carried out.Construct blade structure feature set from the blade root sound signal after time alignment, construct tower cylinder aerodynamic feature set from tower cylinder sound signal.This method uses the self-attention mechanism of Transformer network to fuse blade structure feature set and tower cylinder aerodynamic feature set to generate joint feature vector;Convert structural anomaly probability by calculating reconstruction error through memory-enhanced deep autoencoder.Synchronously analyze the amplitude, pulse width and cross-correlation consistency of sweep tower pulse, generate comprehensive consistency index and convert into sweep tower anomaly probability.Finally, based on real-time power ratio, the above two types of anomaly probability are dynamically weighted and fused to generate comprehensive anomaly probability to determine the blade state.This method deeply fuses multi-source acoustic signals and operating data, uses deep feature extraction and adaptive fusion technology, which can accurately identify blade structural damage and aerodynamic anomaly, and provides technical support for wind turbine blade preventive maintenance.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wind power generation equipment state monitoring, in particular to the technology of monitoring the structural health and aerodynamic performance of wind turbine blades. Specifically, the present application provides a wind turbine blade acoustic dual-path anomaly detection method and system, which realizes early, accurate and robust detection of potential damage and aerodynamic anomalies of the blade, to solve the technical problems of large monitoring blind area, strong noise interference and low detection accuracy existing in the existing single-point acoustic monitoring method. BACKGROUND

[0002] Wind power generation, as an important part of clean energy, occupies an increasingly important position in the global energy structure. Wind turbine blades, as the core components for capturing wind energy, directly affect the power generation efficiency and equipment safety. During long-term operation, blades are susceptible to structural defects such as cracks, delamination, and surface damage due to factors such as fatigue load, environmental corrosion, and lightning strikes. Meanwhile, aerodynamic performance may also decline due to factors such as surface contamination and icing. Wind turbine blades are critical components of wind energy conversion systems, and their performance directly determines the power generation efficiency and economic benefits of wind turbines. Modern wind turbine blades are primarily made of advanced composite materials, including glass fiber reinforced plastic and carbon fiber reinforced plastic, which offer high specific strength, corrosion resistance, and design flexibility. However, with the large-scale development of blade size and the complexity of operating environments, blades face unprecedented technical challenges during service.

[0003] According to international wind power operation data statistics, blade failures account for about 15-20% of total wind turbine failures. Although the failure frequency is relatively low, blade failures often result in long-term downtime and high maintenance costs, having a significant impact on the economic benefits of wind farms. The main problems brought about by the large-scale development of blades include: a sharp increase in structural load, prominent material fatigue issues, increased manufacturing process complexity, and greater difficulty in transportation and installation. Meanwhile, blades operate in harsh natural environments for long periods, needing to withstand wind loads, temperature changes, ultraviolet radiation, rain, snow, icing, lightning strikes, and other environmental factors, all of which pose a severe challenge to the long-term reliability of blades.

[0004] At present, wind turbine blade condition monitoring mainly adopts vibration monitoring, strain monitoring, optical fiber sensing and other methods. Vibration monitoring mainly targets the hub and transmission system, and has limited monitoring effect on the blade itself; strain monitoring requires a large number of sensors to be arranged on the surface of the blade, which is difficult to install and maintain; the optical fiber sensing technology has high cost and is easily affected by environmental factors. Acoustic monitoring, as a new emerging non-contact monitoring technology, has the advantages of easy installation, low cost and wide monitoring range. It has been widely used in the field of mechanical equipment fault diagnosis. Its basic principle is to analyze the sound signals generated during the operation of the equipment, extract characteristic information reflecting the state of the equipment, and then identify the abnormal state of the equipment. In the field of wind power, acoustic monitoring technology is mainly applied to the fault diagnosis of gearboxes, generators and other transmission systems, and good results have been achieved. However, the existing acoustic monitoring methods mainly have the following problems: single-point monitoring cannot fully reflect the overall state of the blade, and blind spots are prone to occur; environmental noise interference is serious, especially the influence of wind noise and mechanical noise; there is a lack of effective feature extraction and fusion methods, making it difficult to accurately identify abnormal states; there is a lack of time-space alignment methods considering the rotating characteristics of the blade; the accuracy and real-time performance of the abnormal detection algorithm need to be improved.

[0005] Therefore, it is urgent to develop a wind turbine blade acoustic abnormality detection method that can overcome the above problems. SUMMARY

[0006] The present application aims to research a wind turbine blade acoustic dual-path abnormality detection method and system to solve the technical problems of monitoring blind area, noise interference, insufficient feature extraction, difficulty in time-space alignment and low detection accuracy in the prior art.

[0007] In a first aspect, the present application provides a wind turbine blade acoustic dual-path abnormality detection method, which comprises:

[0008] Step S1, respectively collecting blade acoustic signals at the blade root measuring point and blade acoustic signals at the tower measuring point through acoustic sensors arranged at the root of each generator blade and an acoustic sensor array arranged in a ring shape at the same height level of the tower, and screening effective data from the blade acoustic signals at the blade root measuring point and the blade acoustic signals at the tower measuring point according to power division working conditions;

[0009] Step S2, based on the collected real-time wind speed and turbulence intensity, using dynamic spectrum subtraction to filter the background noise of the screened blade acoustic signals at the blade root measuring point and the blade acoustic signals at the tower measuring point, and time-aligning the filtered blade acoustic signals at the blade root measuring point and the blade acoustic signals at the tower measuring point;

[0010] Step S3, extracting a blade structure feature set and a tower aerodynamic feature set from the time-aligned blade acoustic signals of the blade root measuring point and the blade acoustic signals of the tower measuring point respectively, inputting the blade structure feature set and the tower aerodynamic feature set into a Transformer model to obtain a joint feature vector;

[0011] Step S4, inputting the joint feature vector into a memory-enhanced deep autoencoder to calculate a reconstruction error, and converting the reconstruction error into a structural anomaly probability output through a probability conversion function based on a historical error distribution;

[0012] Step S5, performing envelope detection on the filtered blade acoustic signals of the tower measuring point to extract a time-domain pulse envelope, obtaining a comprehensive consistency index through consistency analysis on a plurality of continuous pulses, comparing the comprehensive consistency index with a health benchmark model to generate a deviation score, and converting the deviation score into a tower scanning anomaly probability output through a probability model;

[0013] Step S6, adaptively weighting and fusing the structural anomaly probability and the tower scanning anomaly probability to obtain a comprehensive anomaly probability, and determining that the wind turbine blade acoustic anomaly occurs when the comprehensive anomaly probability exceeds an anomaly probability threshold.

[0014] As a preferred embodiment, the background noise filtering processing of the screened blade acoustic signals of the blade root measuring point and the blade acoustic signals of the tower measuring point includes:

[0015] The blade acoustic signals of the blade root measuring point and the blade acoustic signals of the tower measuring point are taken as noisy signals;

[0016] According to the real-time wind speed and turbulence intensity parameters, the dynamic noise power spectrum is obtained through bilinear interpolation of the pre-stored noise model;

[0017] Based on the dynamic noise power spectrum, the instantaneous signal-to-noise ratio of each frequency point of the frequency spectrum of the noisy signal is calculated, and the spectral subtraction factor is adjusted in real time according to the instantaneous signal-to-noise ratio;

[0018] The scaling value of the dynamic noise power spectrum is calculated according to the spectral subtraction factor and the dynamic noise power spectrum;

[0019] The scaling value of the dynamic noise power spectrum is subtracted from the noisy signal to perform background noise filtering processing.

[0020] As a preferred embodiment, the time alignment of the filtered blade acoustic signals of the blade root measuring point and the blade acoustic signals of the tower measuring point includes:

[0021] The blade position is detected, the amplitude change of the blade acoustic signals of the tower measuring point is analyzed, and the blade passing time is determined through acoustic signal peak detection:

[0022] Based on the detected blade passing time, the optimal time delay of the blade root measurement point blade acoustic signal and the tower cylinder measurement point blade acoustic signal is calculated by using a sliding window cross-correlation function;

[0023] Considering the influence of wind speed change on sound propagation, the optimal time delay is dynamically compensated to obtain a compensated time delay to establish a time synchronization relationship between the blade interior and the tower cylinder signal:

[0024] Based on the time synchronization relationship between the blade interior and the tower cylinder signal, the filtered blade root measurement point blade acoustic signal and the tower cylinder measurement point blade acoustic signal are time-aligned.

[0025] As a preferred embodiment, the blade structure feature set and the tower cylinder aerodynamic feature set are extracted from the time-aligned blade root measurement point blade acoustic signal and the tower cylinder measurement point blade acoustic signal, respectively, including:

[0026] The time-aligned blade root measurement point blade acoustic signal is feature-extracted to obtain a blade structure feature set containing multi-scale spectral scatter features, cepstrum envelope coefficients, mel-frequency cepstrum coefficients, and their first and second derivatives;

[0027] The time-aligned tower cylinder measurement point blade acoustic signal is feature-extracted, and its statistical moment features are calculated based on the envelope signal obtained by Hilbert transform demodulation, and further, the spectral centroid and spectral flatness features of the envelope spectrum are extracted to form a tower cylinder aerodynamic feature set.

[0028] As a preferred embodiment, inputting the blade structure feature set and the tower cylinder aerodynamic feature set into a Transformer model to obtain a joint feature vector includes:

[0029] The blade structure feature set and the tower cylinder aerodynamic feature set are adjusted to the same dimension by zero padding to obtain a dimension-unified feature sequence, and the dimension-unified feature sequence is spliced along the dimension direction to obtain a feature matrix;

[0030] A learnable CLS identifier vector is inserted into the first column of the feature matrix to obtain an expanded feature matrix;

[0031] The expanded feature matrix is input into a Transformer model, and the features at each position in the matrix are associated and calculated by a multi-head self-attention mechanism to output a CLS vector that integrates all feature information as a joint feature vector.

[0032] As a preferred embodiment, inputting the joint feature vector into a memory-enhanced deep autoencoder to calculate a reconstruction error includes:

[0033] inputting the joint feature vector into an encoder of the MemAE to obtain a latent code, and calculating a cosine similarity between the latent code and a normal working condition prototype vector stored in a memory module;

[0034] generating an attention weight of each normal working condition prototype vector according to the cosine similarity;

[0035] performing weighted aggregation on the prototype vectors in the memory module by using the attention weight;

[0036] inputting the vector after the weighted aggregation into a decoder of the MemAE to perform feature reconstruction, and calculating a mean square error between the joint feature vector and a vector output by the decoder as the reconstruction error.

[0037] As a preferred embodiment, converting the reconstruction error into a structural anomaly probability output by a probability conversion function based on a historical error distribution includes:

[0038] calculating a structural anomaly probability corresponding to the current reconstruction error by using a Gaussian distribution cumulative distribution function of the reconstruction error of the historical samples in the healthy running state, and the structural anomaly probability is equal to 1 minus the value of the current reconstruction error in the Gaussian distribution cumulative distribution function.

[0039] As a preferred embodiment,

[0040] obtaining a comprehensive consistency index by consistency analysis on the continuous multiple pulses includes taking three continuous pulses as an analysis unit; in one analysis unit, the amplitude consistency of the three pulses and the pulse width consistency of the three pulses are calculated respectively, and the time domain cross-correlation index between the three pulses is calculated in the analysis unit; and the comprehensive consistency index is obtained by fusing the amplitude consistency, the pulse width consistency and the time domain cross-correlation index;

[0041] comparing the comprehensive consistency index with a health benchmark model to generate a deviation score, converting the deviation score into a tower abnormality probability output by a probability model, and mapping the deviation score to the tower abnormality probability by using a Sigmoid function, wherein the steepness parameter and the center point threshold parameter of the Sigmoid function are determined by fitting the distribution characteristics of the deviation score in the healthy running data.

[0042] The construction of the health benchmark model includes: collecting the tower pulse signals without failure alarm for continuous preset days, calculating the mean and standard deviation of the comprehensive consistency index in a fixed time window; taking an interval formed by the mean plus or minus three times the standard deviation as a health benchmark interval, and mapping the deviation score exceeding the interval to the interval of 0 to 1 according to a preset rule.

[0043] As a preferred embodiment, the weight distribution rule in the adaptive weighted fusion of the structural anomaly probability and the tower anomaly probability to obtain the comprehensive anomaly probability comprises: dynamically distributing the weight according to the proportion of the real-time power of the wind turbine to the rated power corresponding to the blade acoustic signal of the blade root measuring point and the blade acoustic signal of the tower measuring point, the higher the proportion of the real-time power, the greater the weight given to the structural anomaly probability, and the weight given to the tower anomaly probability decreases accordingly; the lower the proportion of the real-time power, the smaller the weight given to the structural anomaly probability, and the weight given to the tower anomaly probability increases accordingly.

[0044] In a second aspect, the present application provides a wind turbine blade acoustic dual-path anomaly detection system, comprising:

[0045] A collection module is configured to collect blade acoustic signals of blade root measuring points and blade acoustic signals of tower measuring points by using acoustic sensors arranged at the root of each generator blade and an acoustic sensor array arranged in a ring at the same height level of the tower, and to filter effective data from the blade acoustic signals of blade root measuring points and the blade acoustic signals of tower measuring points according to power division working conditions.

[0046] A processing module is configured to perform background noise filtering processing on the filtered blade acoustic signals of blade root measuring points and the blade acoustic signals of tower measuring points based on the collected real-time wind speed and turbulence intensity, and to perform time alignment on the filtered blade acoustic signals of blade root measuring points and the blade acoustic signals of tower measuring points.

[0047] A model processing module is configured to extract a blade structure feature set and a tower aerodynamic feature set from the time-aligned blade acoustic signals of blade root measuring points and the blade acoustic signals of tower measuring points, respectively, and to input the blade structure feature set and the tower aerodynamic feature set into a Transformer model to obtain a joint feature vector.

[0048] A structural anomaly probability output module is configured to input the joint feature vector into a memory-enhanced deep autoencoder to calculate a reconstruction error, and to convert the reconstruction error into a structural anomaly probability output through a probability conversion function based on historical error distribution.

[0049] A tower anomaly probability output module is configured to perform envelope detection to extract a time-domain pulse envelope from the filtered blade acoustic signals of tower measuring points, to obtain a comprehensive consistency index through consistency analysis on a plurality of consecutive pulses, to compare the comprehensive consistency index with a health benchmark model to generate a deviation score, and to convert the deviation score into a tower anomaly probability output through a probability model.

[0050] A determination module is configured to adaptively and weightedly fuse the structural anomaly probability and the tower anomaly probability to obtain a comprehensive anomaly probability, and to determine that there is an acoustic anomaly of a wind turbine blade when the comprehensive anomaly probability exceeds an anomaly probability threshold.

[0051] Compared with the prior art, the present application has the following beneficial effects: the present application deeply fuses multi-source acoustic signals and operation data, uses deep feature extraction and adaptive fusion technology, can accurately identify blade structure damage and aerodynamic abnormalities, and provides technical support for preventive maintenance of wind turbine blades. BRIEF DESCRIPTION OF DRAWINGS

[0052] In order to more clearly illustrate the technical solutions implemented by the present application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings described below are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0053] Figure 1 Flow chart of the wind turbine blade acoustic dual-path anomaly detection method of the present application;

[0054] Figure 2 Flow chart of the sound signal filtering and time series alignment processing of the present application;

[0055] Figure 3 Flow chart of the structural anomaly detection of the present application. DETAILED DESCRIPTION

[0056] The present application will be further described below in conjunction with the drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present application, and cannot limit the protection scope of the present application.

[0057] The present embodiment provides a specific implementation scheme of a wind turbine blade acoustic anomaly detection method, as shown in Figure 1 The method comprises the following steps: data acquisition and filtering, dynamic spectral subtraction background noise reduction, sound signal time synchronization and feature extraction, Transformer feature fusion, MemAE reconstruction error calculation, structural anomaly probability conversion, time domain envelope pulse consistency analysis, sweep tower anomaly probability conversion, comprehensive anomaly probability fusion, and blade acoustic anomaly identification. Specifically:

[0058] S1: through three acoustic sensor arrays arranged at 120° in a ring shape at the same height level of the roots of three wind turbine blades and the tower drum, synchronously collect multi-channel sound signals, specifically including blade acoustic signals at blade root measuring points and blade acoustic signals at tower drum measuring points; based on the real-time power of the wind turbine corresponding to the sound signals, filter the sound signals in the real-time power range of 10%-90% of the rated power;

[0059] S2: Establish a noise model driven by wind speed and turbulence intensity, filter the cleaned data using dynamic spectral subtraction, and perform acoustic signal peak detection and dynamic time delay compensation on the filtered signal to align the data on all sensor signal time series;

[0060] S3: Extract a blade structure feature set from the blade root sound signal after time alignment in S2, and extract a tower drum aerodynamic feature set from the tower drum sound signal; input the blade structure feature set and the tower drum aerodynamic feature set into a Transformer model, and use its self-attention mechanism for multi-dimensional feature fusion to obtain a joint feature vector;

[0061] S4: Input the joint feature vector output by S3 into a memory-enhanced deep autoencoder (MemAE); the MemAE calculates a reconstruction error based on its encoder, memory module, and decoder, and converts the reconstruction error into a structural anomaly probability output through a probability conversion function based on historical error distribution;

[0062] S5: Envelope detection is performed on the tower acoustic signal processed in S2, and the time-domain pulse envelope of each sweep tower period is extracted. Taking three consecutive pulses as a unit, the amplitude, pulse width, and cross-correlation consistency index are weighted and fused to generate a comprehensive consistency index. The deviation score is compared with the healthy benchmark model to generate a sweep tower anomaly probability output;

[0063] S6: Construct a dynamic weight distribution module to generate adaptive weights for the structural anomaly probability and the sweep tower anomaly probability based on real-time power data. The comprehensive anomaly probability is obtained by weighted fusion. When the comprehensive anomaly probability exceeds the dynamically adjusted anomaly probability threshold, it is determined that there is an acoustic anomaly in the wind turbine blade.

[0064] Each step is described in detail as follows:

[0065] In S1, acoustic sensors are installed on the inner roots of the three blades. The sensors are directed towards the blade tip and are located at the blade root to collect acoustic signals generated during the operation of the blade internal structure, i.e., blade acoustic signals at the blade root. At the bottom of the tower drum near the blade tip, three acoustic sensors are evenly distributed in a ring at an interval of 120°, which are used to collect aerodynamic noise signals generated during the rotation of the blade, i.e., blade acoustic signals at the tower drum. On this basis, the real-time power of the wind turbine at the corresponding time of the collected sound signal is obtained, and the sound signal in the 10%-90% rated power operating range is selected.

[0066] S2, as Figure 2The sound signal obtained in S1 is shown, and dynamic spectral subtraction is used for noise suppression according to the complex noise characteristics in the wind turbine environment. First, the noise power spectrum is obtained, and according to the real-time wind speed and turbulence intensity parameters, the dynamic noise power spectrum P n (k) is obtained by bilinear interpolation of the pre-stored noise model. Then, the instantaneous signal-to-noise ratio is calculated for each frequency point:

[0067] SNR(k)=10log 10 [|Y(k) 2 / P n (k)|]

[0068] Where Y(k) is the frequency spectrum of the noisy signal.

[0069] The spectral subtraction factor is adjusted according to the real-time signal-to-noise ratio:

[0070] α(k)=α0(k)·[1+γ / SNR(k)]

[0071] Finally, the noisy signal spectrum is processed by spectral subtraction:

[0072]

[0073] To avoid excessive suppression, set the minimum gain limit:

[0074]

[0075] The sound data after filtering is processed in time scale, the peak value of acoustic signal detection is used to determine the blade passing time, the sliding window cross-correlation function is used to calculate the optimal time delay, and dynamic compensation is performed to establish the time synchronization relationship between the blade internal and tower signals. Specifically, first, the blade position is detected, the amplitude change of the tower sensor signal is analyzed, and the blade passing time is detected:

[0076]

[0077] Where t j (t) is the tower sound signal, and H[·] is the Hilbert transform.

[0078] Then, the time delay estimation is performed, and the sliding window cross-correlation function is used to calculate the optimal time delay between the blade internal signal and the tower signal:

[0079]

[0080] Considering the influence of wind speed change on sound propagation, dynamic compensation of time delay is performed:

[0081]

[0082] Where, is the position vector from the blade sensor to the tower sensor, is the wind speed vector, c0is the speed of sound in air, is the time delay change caused by the wind speed.

[0083] Finally, the data alignment of the two sound signals in the time sequence is realized:

[0084]

[0085] S3, multi-dimensional feature extraction and feature fusion are performed on the processed blade sound data, blade structure feature sets are extracted from the blade sound signals of the blade internal measuring points, and tower aerodynamic feature sets are extracted from the blade sound signals of the tower measuring points based on Hilbert transform demodulation. Then, the blade structure features and the tower aerodynamic feature sets are input into the Transformer for feature fusion. The specific steps are as follows: first, feature extraction is performed on the blade root sound signals that have been time-aligned in S2 to obtain blade structure feature sets containing multi-scale spectral dispersion features, cepstrum envelope coefficients, mel-frequency cepstrum coefficients, and first and second derivatives thereof. At the same time, feature extraction is performed on the blade sound signals of the tower measuring points, the statistical moment features of the envelope signals obtained based on Hilbert transform demodulation are calculated, and the spectral centroid and spectral flatness features of the envelope spectrum are further extracted to form the tower aerodynamic feature sets. Since the lengths of the two feature sets are different, the shorter feature sequence is padded to the same dimension as the longer feature sequence. After obtaining the feature sequences with uniform dimensions, the two are spliced in the feature dimension direction to form a feature matrix M. Then, a learnable CLS identification vector is inserted into the first column of the feature matrix M, and the dimension of the vector is the same as that of the feature matrix M. At this time, the feature matrix is expanded to M'; the input matrix is input into the Transformer model, and the multi-head self-attention mechanism

[0086]

[0087] The features at each position in the matrix are calculated to realize deep fusion of the multi-dimensional features, and the CLS identification vector fully aggregates the global information of the blade structure feature sets and the tower aerodynamic feature sets. After output from the Transformer model, the CLS vector that has fused all the feature information is selected as the joint feature vector.

[0088] S4, as Figure 3 shown, the joint feature vector output by S3 is input into the memory-enhanced deep autoencoder MemAE. The model includes an encoder, a memory module, and a decoder. First, the encoder processes the input joint feature vector to obtain a latent code. Then, the cosine similarity between the latent code and the normal working condition prototype vector stored in the memory module is calculated. The formula for calculating the cosine similarity is:

[0089]

[0090] wherein sim i is the similarity of z and the i-th memory item; z is the latent code; m i is the memory item.

[0091] Then, an attention weight of each normal operating mode prototype vector is generated according to the cosine similarity:

[0092]

[0093] The prototype vectors in the memory module are weighted and aggregated using the attention weight:

[0094]

[0095] wherein z' is a new latent vector of the aggregation output.

[0096] After that, the new latent vector z after weighted aggregation is input into the decoder of the MemAE for feature reconstruction, and the mean square error between the joint feature vector and the decoder reconstruction output vector is calculated as the reconstruction error, and the calculation formula of the mean square error is:

[0097]

[0098] wherein n is the vector dimension, y i is the i-th element of the joint feature vector, y' i is the i-th element of the decoder reconstruction output vector; after obtaining the reconstruction error MSE of the current sample, it needs to be converted into a structural anomaly probability representing the abnormal possibility. First, the mean μ and the standard deviation σ of the reconstruction error of a large number of historical samples in the system health running state are calculated, and a Gaussian distribution model N(μ,σ 2 ) is established. Then, the cumulative distribution function CDF of the distribution is calculated, that is, the probability that the reconstruction error is less than or equal to MSE under the health state. Finally, the structural anomaly probability P1 is calculated by the formula

[0099] P1=1-CDF(MSE)

[0100] The structural anomaly probability is calculated, and the greater the value, the higher the possibility of the current state being abnormal.

[0101] S5, envelope detection is performed on the tower drum acoustic signal after S2 preprocessing to highlight the pulse feature, and the time domain pulse waveform corresponding to each period is accurately extracted based on the blade passing period. In order to evaluate the dynamic consistency between pulses, three consecutive tower scanning pulses are divided into a basic analysis unit. Within each analysis unit, three key sub-indices are calculated in parallel: amplitude consistency C aAmplitude fluctuation C ω Pulse width consistency at a specific threshold C x The normalized cross-correlation coefficient between each pair of pulses in the unit is calculated, and the maximum or average value of the three pairs is taken as the index value to measure the waveform time domain similarity. Then, the three sub-indices are summed by a weighted sum formula:

[0102] I c = w a · C a + w ω · C ω + w x · C x

[0103] Fusion into a comprehensive consistency index I c , where the weight coefficients w a , w ω , w x are dynamically configured according to the calibration data of the deployed acoustic sensor. The calibration process analyzes the sensitivity, stability of each sub-index to disturbance and its contribution to the overall consistency under healthy state, and determines the optimal weight combination through data analysis or expert experience combined optimization, and the sum of the weights is 1. To quantify the abnormality of I c , a health benchmark model needs to be built: collect the tower scanning pulse signals within the pre-set number of days of continuous fault-free alarm of the fan, calculate the sample mean μ and sample standard deviation σ of all I c values in the fixed time window, and define the health benchmark interval [μ-3σ, μ+3σ] based on the normal distribution assumption. For the real-time calculated I c , calculate its deviation score: if I c is within the health interval, the value of s is 0; if it is below the lower limit, the value of s is ; if it is above the upper limit, the value of s is Ensure that the deviation score is linearly mapped to the interval [0, 1], and the larger the value, the farther it deviates from the health benchmark. Finally, input the deviation score into the Sigmoid probability conversion function:

[0104]

[0105] Calculate the tower scanning abnormality probability P2. The key parameters of the function, the steepness parameter k and the center point threshold s0, are determined by fitting the distribution characteristics of the deviation score in the health running data set: this process finally realizes the accurate quantification of the tower scanning pulse consistency and converts it into an intuitive, probability-based abnormal risk index, which can be directly used for early warning or as input for advanced diagnostic systems.

[0106] In the dynamic weight allocation module of step S6, adaptive weights are generated based on real-time operating power data, specifically: calculating the real-time power P as a percentage of the wind turbine's rated power P. rated The ratio r = P / P rated Based on this, the weight w1 = r of the structural anomaly probability is determined. 2 The weights of the abnormal probability of the tower sweeping are w2 = 1 - r 2 The quadratic function design rule ensures the dynamic adaptability of the weights. When the wind turbine is operating at high power, w1 increases significantly while w2 decreases accordingly. The comprehensive anomaly probability determination focuses more on the generalized anomaly detection capability of the MemAE model. When the wind turbine is operating at low power, w1 decreases significantly while w2 increases accordingly. The comprehensive anomaly probability determination focuses more on the physical consistency analysis of the tower sweep pulse signal. Thus, the advantages of the two detection methods are optimized and integrated according to the characteristics of different power ranges, providing a basis for generating a comprehensive anomaly probability that accurately reflects the blade state.

[0107] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A wind turbine blade acoustic dual-path anomaly detection method, characterized in that, The method comprises the following steps: Step S1, respectively collecting blade acoustic signals at the blade root measuring point and blade acoustic signals at the tower measuring point through acoustic sensors arranged at the root of each generator blade and an acoustic sensor array arranged in a ring at the same height level of the tower drum, and screening effective data from the blade acoustic signals at the blade root measuring point and the blade acoustic signals at the tower measuring point according to power division working conditions; Step S2, based on the collected real-time wind speed and turbulence intensity, performing background noise filtering processing on the screened blade acoustic signals at the blade root measuring point and the blade acoustic signals at the tower measuring point by using dynamic spectral subtraction, and performing time alignment on the blade acoustic signals at the blade root measuring point and the blade acoustic signals at the tower measuring point after filtering processing; Step S3, extracting a blade structure feature set and a tower aerodynamic feature set from the time-aligned blade acoustic signals at the blade root measuring point and the blade acoustic signals at the tower measuring point, respectively, and inputting the blade structure feature set and the tower aerodynamic feature set into a Transformer model to obtain a joint feature vector; Step S4, inputting the joint feature vector into a memory-enhanced deep autoencoder to calculate a reconstruction error, and converting the reconstruction error into a structural anomaly probability output through a probability conversion function based on historical error distribution; Step S5, performing envelope detection on the filtered blade acoustic signals at the tower measuring point to extract a time-domain pulse envelope, obtaining a comprehensive consistency index through consistency analysis on a plurality of continuous pulses, comparing the comprehensive consistency index with a health benchmark model to generate a deviation score, and converting the deviation score into a tower scanning anomaly probability output through a probability model; Step S6, adaptively weighting and fusing the structural anomaly probability and the tower scanning anomaly probability to obtain a comprehensive anomaly probability, and determining that there is an acoustic anomaly of the wind power blade when the comprehensive anomaly probability exceeds an anomaly probability threshold.

2. A wind generator blade acoustic dual-path anomaly detection method according to claim 1, characterised in that, The background noise filtering processing of the screened blade acoustic signals at the blade root measuring point and the blade acoustic signals at the tower measuring point by using dynamic spectral subtraction comprises: The blade acoustic signals at the blade root measuring point and the blade acoustic signals at the tower measuring point are taken as noise-containing signals; According to the real-time wind speed and turbulence intensity parameters, a dynamic noise power spectrum is obtained through bilinear interpolation of a pre-stored noise model; Based on the dynamic noise power spectrum, an instantaneous signal-to-noise ratio of each frequency point of the frequency spectrum of the noise-containing signal is calculated, and a spectral subtraction factor is adjusted in real time according to the instantaneous signal-to-noise ratio; The scaling value of the dynamic noise power spectrum is calculated according to the spectral subtraction factor and the dynamic noise power spectrum; The scaling value of the dynamic noise power spectrum is subtracted from the noise-containing signal to perform background noise filtering processing.

3. A wind generator blade acoustic dual-path anomaly detection method according to claim 1, characterized in that, The time alignment of the filtered blade acoustic signals at the blade root measuring point and the blade acoustic signals at the tower measuring point comprises: The blade position is detected, the amplitude change of the blade acoustic signals at the tower measuring point is analyzed, and the blade passing time is determined through acoustic signal peak detection; Based on the detected blade passing time, time delay estimation is performed, and the optimal time delay of the blade acoustic signals at the blade root measuring point and the blade acoustic signals at the tower measuring point is calculated by using a sliding window cross-correlation function; Considering the influence of wind speed change on sound propagation, the optimal time delay is dynamically compensated to obtain a compensated time delay, so as to establish the time synchronization relationship between the blade inside and the tower signals: Time aligning the filtered blade root point blade acoustic signal and the tower point blade acoustic signal based on the time synchronization relationship between the blade interior and the tower signal.

4. A wind generator blade acoustic dual-path anomaly detection method according to claim 1, characterized in that, Extracting a blade structure feature set and a tower aerodynamic feature set from the time-aligned blade root point blade acoustic signal and the tower point blade acoustic signal respectively includes: Performing feature extraction on the time-aligned blade root point blade acoustic signal to obtain the blade structure feature set including a multi-scale spectrum divergence feature, a cepstrum envelope coefficient, a mel-frequency cepstrum coefficient, and first and second derivatives thereof; Performing feature extraction on the time-aligned tower point blade acoustic signal, calculating statistical moment features based on an envelope signal obtained by Hilbert transform demodulation, and further extracting spectral centroid and spectral flatness features of the envelope spectrum to form the tower aerodynamic feature set.

5. A wind generator blade acoustic dual-path anomaly detection method according to claim 1, characterized in that, Inputting the blade structure feature set and the tower aerodynamic feature set into a Transformer model to obtain a joint feature vector includes: Adjusting the blade structure feature set and the tower aerodynamic feature set to the same dimension in a zero-padding manner to obtain a dimension-unified feature sequence, and concatenating the dimension-unified feature sequence along the dimension direction to obtain a feature matrix; Inserting a learnable CLS identifier vector into the first column of the feature matrix to obtain an expanded feature matrix; Inputting the expanded feature matrix into the Transformer model, and performing correlation calculation on the features at each position in the matrix through a multi-head self-attention mechanism to take the output CLS vector fused with all feature information as the joint feature vector.

6. A wind generator blade acoustic dual-path anomaly detection method according to claim 1, characterized in that, Inputting the joint feature vector into a memory-enhanced deep autoencoder to calculate a reconstruction error includes: Inputting the joint feature vector into the encoder of the MemAE to obtain a latent code, and calculating a cosine similarity between the latent code and a normal working condition prototype vector stored in the memory module; Generating an attention weight for each normal working condition prototype vector according to the cosine similarity; Weighted aggregation of the prototype vectors in the memory module using the attention weight; Inputting the weighted aggregated vector into the decoder of the MemAE for feature reconstruction, and calculating a mean square error between the joint feature vector and the decoder reconstruction output vector as the reconstruction error.

7. A wind generator blade acoustic dual-path anomaly detection method according to claim 1, characterized in that, Converting the reconstruction error into a structural anomaly probability output through a probability conversion function based on a historical error distribution includes: Using a Gaussian distribution cumulative distribution function of historical sample reconstruction errors in a healthy running state to calculate a structural anomaly probability corresponding to the current reconstruction error, and the structural anomaly probability is equal to 1 minus the value of the current reconstruction error in the Gaussian distribution cumulative distribution function.

8. The wind turbine blade acoustic dual-path anomaly detection method according to claim 1, wherein: Obtaining a comprehensive consistency index through consistency analysis on a plurality of continuous pulses includes taking three consecutive pulses as an analysis unit; in an analysis unit, the amplitude consistency of the three pulses and the pulse width consistency of the three pulses are calculated respectively, and the time domain cross-correlation index between the three pulses is calculated in the analysis unit; and the amplitude consistency, the pulse width consistency, and the time domain cross-correlation index are fused to obtain the comprehensive consistency index. The comprehensive consistency index is compared with a health benchmark model and a deviation score is generated, the deviation score is converted into a tower abnormality probability output by a probability model, and the deviation score is mapped into a tower abnormality probability using a Sigmoid function, wherein a steepness parameter and a center point threshold parameter of the Sigmoid function are determined by fitting a distribution characteristic of the deviation score in health operation data; The construction of the health benchmark model comprises: collecting a sweep tower pulse signal without failure alarm for a continuous preset number of days, and calculating the mean and standard deviation of the comprehensive consistency index in a fixed time window; An interval formed by the mean plus or minus three times the standard deviation is taken as a health benchmark interval, and a deviation score exceeding the interval is mapped to an interval of 0 to 1 according to a preset rule.

9. The wind turbine blade acoustic dual-path anomaly detection method of claim 1, wherein, The weight distribution rule in adaptive weighted fusion of the structural abnormality probability and the tower abnormality probability to obtain a comprehensive abnormality probability comprises: dynamically distributing weights according to the proportion of real-time power of the wind turbine to rated power corresponding to the blade acoustic signal at the blade root measuring point and the blade acoustic signal at the tower measuring point, the higher the real-time power proportion, the greater the weight given to the structural abnormality probability, and the weight given to the tower abnormality probability is correspondingly reduced; the lower the real-time power proportion, the smaller the weight given to the structural abnormality probability, and the weight given to the tower abnormality probability is correspondingly increased.

10. A wind turbine blade acoustic dual-path anomaly detection system, characterized in that, Comprise: The acquisition module is configured to acquire blade acoustic signals at blade root measuring points and blade acoustic signals at tower measuring points by acoustic sensors arranged at the roots of respective generator blades and an acoustic sensor array arranged annularly at the same height level of the tower, and to filter out effective data from the blade acoustic signals at the blade root measuring points and the blade acoustic signals at the tower measuring points according to power division working conditions; The processing module is configured to perform background noise filtering processing on the filtered blade acoustic signals at the blade root measuring points and the blade acoustic signals at the tower measuring points by using dynamic spectral subtraction based on the acquired real-time wind speed and turbulence intensity, and to perform time alignment on the filtered blade acoustic signals at the blade root measuring points and the blade acoustic signals at the tower measuring points; The model processing module is configured to extract a blade structure feature set and a tower aerodynamic feature set from the time-aligned blade acoustic signals at the blade root measuring points and the blade acoustic signals at the tower measuring points, respectively, and to input the blade structure feature set and the tower aerodynamic feature set into a Transformer model to obtain a joint feature vector; The structural abnormality probability output module is configured to input the joint feature vector into a memory-enhanced deep autoencoder to calculate a reconstruction error, and to convert the reconstruction error into a structural abnormality probability output by a probability conversion function based on a historical error distribution; The tower abnormality probability output module is configured to perform envelope detection on the filtered blade acoustic signals at the tower measuring points to extract time-domain pulse envelopes, to obtain a comprehensive consistency index by consistency analysis on a plurality of continuous pulses, to compare the comprehensive consistency index with a health benchmark model and generate a deviation score, and to convert the deviation score into a tower abnormality probability output by a probability model. A determination module is configured to adaptively weight and fuse the structural anomaly probability and the scanning tower anomaly probability to obtain a comprehensive anomaly probability, and determine that the wind power blade has an acoustic anomaly when the comprehensive anomaly probability exceeds an anomaly probability threshold.

Images