CN-121976923-A - Voiceprint detection method for internal abnormality of wind driven generator blade

CN121976923ACN 121976923 ACN121976923 ACN 121976923ACN-121976923-A

Abstract

The invention discloses a voiceprint detection method for internal abnormality of a wind driven generator blade, and relates to the technical field of wind driven generator fault detection, comprising the following components of a signal acquisition step, a signal preprocessing step, a characteristic parameter extraction step, a voiceprint-vibration mode association model construction step and a residual life prediction model training and detection step; according to the invention, the internal abnormal type of the blade is accurately detected, the residual life of the blade can be accurately predicted through a deep belief network, the DBN model is trained through abnormal state data of the full life cycle of the blade, the improved contrast divergence algorithm is adopted to optimize the network weight and bias, the self-adaptive learning rate optimization algorithm is combined, the vibration and stagnation problems in the training process are effectively avoided, in addition, the model output layer adopts an improved linear activation function, the influence of the accumulated running time of the blade on the residual life is considered, and the prediction result is more in accordance with the actual attenuation law.

Inventors

SHEN FUJIA
ZHANG XIUQUAN
PAN YONG
WANG BANGYU
Guo songyu

Assignees

大唐黑龙江新能源开发有限公司

Dates

Publication Date: 20260505
Application Date: 20260127

Claims (9)

1. A voiceprint detection method for internal abnormality of a wind driven generator blade is characterized by comprising the following specific steps: The method comprises the steps of collecting signals, namely arranging voiceprint sensors and vibration sensors at preset monitoring points of blades of a wind driven generator, and collecting voiceprint signals and vibration response signals of the blades in normal operation and different internal abnormal states; the method comprises the steps of signal preprocessing, namely sequentially carrying out denoising, pre-emphasis and framing windowing on an acquired voiceprint signal, wherein the denoising adopts a wavelet threshold denoising algorithm, and sequentially carrying out trend removing and filtering on a vibration response signal, and the filtering adopts a Kalman filtering algorithm; The characteristic parameter extraction step comprises the steps of extracting a Mel frequency cepstrum coefficient, a linear prediction cepstrum coefficient, a spectrum centroid and a spectrum bandwidth from a preprocessed voiceprint signal to be used as a voiceprint characteristic parameter set, wherein the Mel frequency cepstrum coefficient extraction adopts a Mel filter bank; The method comprises the steps of establishing a voiceprint-vibration mode association model, namely taking an extracted voiceprint characteristic parameter set and a vibration mode parameter set as inputs, analyzing the correlation of the two parameters by adopting a Pearson correlation coefficient, and screening out characteristic parameters which are strongly correlated with the internal abnormality of the blade; The method comprises the steps of training and detecting a residual life prediction model, wherein fusion characteristic parameters output by a voiceprint-vibration mode association model are input to a deep belief network, the DBN comprises an input layer, a plurality of layers of limited Boltzmann machine hidden layers and an output layer, the DBN is trained through abnormal state data of a full life cycle of a blade, RBM of each layer is trained through a contrast divergence algorithm, network weight and bias are optimized, a self-adaptive learning rate optimization algorithm is adopted in the training process, voiceprint signals and vibration response signals of the blade to be detected are input after training is completed, and the abnormal type and the corresponding residual life of the blade are output through the DBN.
2. The voiceprint detection method for the internal anomalies of the wind driven generator blade according to claim 1, wherein the denoising processing in the signal preprocessing step adopts a self-adaptive multi-scale wavelet threshold denoising algorithm, after wavelet transformation is carried out on voiceprint signals, denoising thresholds are dynamically adjusted according to wavelet coefficients of different decomposition layers, wherein the standard deviation of noise of each layer is calculated through a median absolute deviation method, namely the median of absolute values of all wavelet coefficients of the layer is divided by 0.6745, the self-adaptive adjustment factor of each layer is determined by the energy duty ratio of the wavelet coefficients of the layer, namely the ratio of the energy of the layer to the total energy of all decomposition layers, and the wavelet coefficients are combined with dynamic thresholds constructed based on the standard deviation and the adjustment factor.
3. The method for detecting abnormal sound patterns in a wind turbine blade according to claim 1, wherein the mel frequency cepstrum coefficient extraction in the characteristic parameter extraction step adopts an improved mel filter bank, and a center frequency calculation formula thereof is as follows: wherein Is the first The center frequency of the individual mel-filters, As a reference to the frequency of the reference, At the highest cut-off frequency of the frequency, For the filter sequence number, For the total number of filters, in the cepstral coefficient calculation, a weighted DCT transform is used: wherein Is the first The number of mel-frequency cepstral coefficients, Is the number of the cepstral coefficient, Is the first The output energy of the individual mel-filters, The weight coefficient is determined by the abnormal correlation of the energy of the frequency band.
4. The voiceprint detection method for internal anomalies of a wind turbine blade according to claim 1, wherein in the feature parameter extraction step, a modified random subspace method is adopted for vibration mode parameter extraction, and a feature value stability criterion formula for mode identification is as follows: wherein Is used as an index of the stability of the characteristic value, For the number of sub-space decompositions, In order to decompose the number of times sequence number, Is the first The characteristic value obtained by the secondary subspace decomposition, Is that The characteristic value average value of the secondary decomposition, Is the standard deviation of the characteristic value, And judging the stable mode, and correcting the natural frequency calculation by combining the structural parameters of the blade: wherein To the modified first The natural frequency of the step is set, In order to identify the natural frequency of the signal, 、 The elastic modulus and the section moment of inertia of the blade at the current moment are respectively, 、 Is an initial state parameter.
5. The voiceprint detection method for internal anomalies of a wind turbine blade according to claim 1, wherein the correlation analysis in the voiceprint-vibration mode correlation model construction step adopts an improved pearson correlation coefficient algorithm, and the calculation formula is: wherein Is characteristic parameter of voiceprint And vibration mode parameters Is used for the improvement of the pearson correlation coefficient, As a total number of samples, For the sample sequence number, 、 Sample values of voiceprint characteristics and vibration mode parameters respectively, 、 Is the mean value of the two values, For sample weight, the screening threshold adopts a dynamic adjustment mechanism: wherein A threshold value is selected for the feature, For the number of abnormal samples, Is the total number of samples.
6. The voiceprint detection method for internal anomalies of a wind turbine blade according to claim 1, wherein the weight calculation formula of the feature attention mechanism module in the voiceprint-vibration mode correlation model construction step is: wherein 、 Respectively fusing weights of voiceprint characteristics and vibration mode parameters, 、 The abnormal sensitivity indexes of the two types of characteristics are respectively, Is a temperature coefficient.
7. The voiceprint detection method for internal anomalies of a wind turbine blade according to claim 1, wherein in the voiceprint-vibration mode correlation model construction step, a correlation model adopts an improved gaussian mixture model, and the probability density function is: wherein For fusing feature vectors Is used to determine the probability density value of (1), For the sequence number of the mixed component, As a function of the gaussian distribution, In order to fuse the feature vectors, In order to mix the number of components, As a result of the mixing coefficient, As a function of the gaussian distribution, 、 The mean vector and covariance matrix are respectively, As a state self-adaptive factor, the occurrence probability of the abnormality is dynamically adjusted by: wherein Is the first The number of samples that are in class anomaly, For the minimum sample number category, the model training adopts an improved EM algorithm, and regularization terms are introduced in the M steps: wherein As a function of the likelihood that the time of day, For the regularization coefficient(s), Is the Frobenius norm.
8. The voiceprint detection method for internal anomalies of a wind turbine blade according to claim 1, wherein the limited boltzmann machine training of the deep belief network in the residual life prediction model training and detection step employs an improved contrast divergence algorithm, and the weight update formula is: wherein Is that Time-of-day limited boltzmann machine visible layer And hidden layer Is used for the connection weight of the (c), To hide layer neurons Is to be activated in the active state of (a), Is that Time of day visible layer And hidden layer Is used for the connection weight of the (c), In order for the rate of learning to be high, 、 The expectations of the data distribution and the reconstruction distribution respectively, The error-adjusting factor is reconstructed and the error-adjusting factor is calculated, For the L2 regularization coefficient, introducing an adaptive slope parameter in the hidden layer activation function: wherein As a function of the sigmoid, 、 To learn parameters, the output layer of the DBN employs an improved linear activation function: wherein As a life decay factor, determined by the cumulative running time of the blade: wherein For the current run-time period of time, For the design life.
9. The voiceprint detection method for the internal anomalies of the wind driven generator blade according to claim 1 is characterized in that the specific implementation mode of the self-adaptive learning rate optimization algorithm in the residual life prediction model training and detection step is that the learning rate is dynamically adjusted by monitoring the change rate of a loss function of adjacent iteration cycles in the DBN training process by taking 0.01 as an initial learning rate, when the decreasing proportion of the loss function exceeds 2% in the previous cycle, the current learning rate is increased to 1.2 times to accelerate convergence, when the increasing proportion of the loss function exceeds 2% in the previous cycle, the current learning rate is reduced to 0.8 times to avoid training oscillation, if the change rate of the loss function is between-2% and 2%, the learning rate is kept unchanged, 10 -5 is set as a lower limit of the learning rate to prevent training stagnation caused by the too small learning rate, and for the time sequence characteristics of a blade life sample, a time weighted loss function is used to calculate a training error, wherein the weight is given according to the running time of the blade corresponding to the sample, and the running time is higher as the sample weight of the running time is close to the designed life.

Description

Voiceprint detection method for internal abnormality of wind driven generator blade Technical Field The invention relates to the technical field of wind driven generator fault detection, in particular to a voiceprint detection method for internal abnormality of a wind driven generator blade. Background With the continuous growth of global energy demand and the transformation of energy structures, wind power generation is widely used worldwide as a clean and renewable energy source, and wind power generator blades are used as key components for capturing wind energy, and the health status of the wind power generator blades directly influences the power generation efficiency and the operation safety, however, due to long-term exposure to severe natural environments, the wind power generator blades are susceptible to various internal anomalies such as cracks, layering, hollows and the like, and if the internal anomalies are not detected and treated in time, the structural integrity and the service life of the blades are seriously threatened, and even catastrophic accidents can be caused. The traditional wind driven generator blade internal anomaly detection methods mainly depend on regular visual inspection and off-line testing, are time-consuming and labor-consuming, and are difficult to find early minor defects, in recent years, although some on-line monitoring methods based on vibration analysis and sound emission technologies are presented, the methods usually only pay attention to signal characteristics of a single dimension, and the inherent connection between voiceprint signals and vibration response signals is ignored, so that the detection accuracy and reliability are limited, and especially in a strong noise environment, the traditional methods are difficult to effectively extract characteristic information closely related to the blade internal anomaly, false alarm and false alarm are easy to cause, in addition, the traditional methods have larger defects in the aspect of residual life prediction, the residual service life of the blade is often not accurately estimated, and the operation and maintenance management is challenging. Aiming at the problems of low detection precision, poor reliability, inaccurate residual life prediction and the like of the traditional method for detecting the internal abnormality of the wind driven generator blade, the development of the voiceprint detection method for the internal abnormality of the wind driven generator blade is particularly important. Disclosure of Invention The invention aims to make up the defects of the prior art, provides a voiceprint detection method for the internal abnormality of the wind driven generator blade, can effectively extract characteristic information closely related to the internal abnormality of the blade through a self-adaptive multi-scale wavelet threshold denoising algorithm and an improved random subspace method, constructs a voiceprint-vibration mode association model, realizes high-precision detection of the internal abnormality type of the blade, and simultaneously utilizes a deep belief network to accurately predict the residual life of the blade, thereby providing scientific basis for operation and maintenance management of a wind driven generator field. The invention provides a voiceprint detection method for internal abnormality of a wind driven generator blade, which aims to solve the technical problems and comprises the following specific steps: the method comprises the steps of collecting signals, namely arranging voiceprint sensors and vibration sensors at preset monitoring points of blades of a wind driven generator, and collecting voiceprint signals and vibration response signals of the blades in normal operation and different internal abnormal states, wherein the internal abnormal states comprise cracks, layering and hollows in the blades; the method comprises the steps of signal preprocessing, namely sequentially carrying out denoising, pre-emphasis and framing windowing on an acquired voiceprint signal, wherein the denoising adopts a wavelet threshold denoising algorithm, and sequentially carrying out trend removing and filtering on a vibration response signal, and the filtering adopts a Kalman filtering algorithm; The characteristic parameter extraction step comprises the steps of extracting a Mel frequency cepstrum coefficient, a linear prediction cepstrum coefficient, a spectrum centroid and a spectrum bandwidth from a preprocessed voiceprint signal to be used as a voiceprint characteristic parameter set, wherein the Mel frequency cepstrum coefficient extraction adopts a Mel filter bank; The method comprises the steps of establishing a voiceprint-vibration mode association model, namely taking an extracted voiceprint characteristic parameter set and a vibration mode parameter set as inputs, analyzing the correlation of the two parameters by adopting a Pearson correlation coefficient, and screening out charact