| The wind industry is growing as technology advances and demand for cleaner energy increases.Wind turbine blade is one of the important parts of high cost of the wind turbine.The maintenance cost of wind turbine blade is increased due to long-term fatigue and bad environmental conditions.Researchers have been studying various blade damage detection techniques,but at present,blade damage monitoring mostly relies on wind farm staff to listen and see regularly.This method has low efficiency and high labor cost.Based on the abnormal sound emitted when the fan blade is damaged,a defect detection technology based on octave range and acoustic characteristics analysis of wavelet packet is proposed in this paper.By using the sound of the fan blade when it is running and rotating to monitor the state of the fan blade,the defect detection of the fan blade can be achieved without contact.In this paper,the main components of the sound signal of the fan blade are analyzed,and the acoustic signal acquisition front end suitable for the outdoor working environment of the wind field is constructed.Acoustic signal acquisition experiments are carried out in different directions of the fan to determine the best position of acoustic signal acquisition.Acoustic signal data samples of fan blades were collected from three wind farms in Jiangxi,Hubei and other places.After data cleaning,the samples were labeled in cooperation with the staff of the wind farm and the acoustic signal data set of fan blades was established.Aiming at the complex noise components of the sound signal of the fan blade,this paper focuses on the noise reduction algorithm of the sound signal.The Butterworth band-pass filter is used to do the preliminary denoising pretreatment for the signal noise reduction,and the low frequency wind noise is removed.Aiming at the shortcoming of Butterworth band-pass filter,a new noise reduction method based on the improvement of the acoustic signal characteristics of fan blades is proposed,which combines the continuous mean square criterion and cross-correlation analysis method to reduce the noise of the acoustic signal.Through the comparison and analysis of the time-frequency diagrams before and after noise reduction,it can be concluded that this method can effectively remove some noises,such as human cough,without losing the defect information in the defect sound signal.Based on the characteristics of acoustic signals in the frequency domain,two feature extraction methods are selected,and the advantages and disadvantages of the two methods are compared.First,the highest classification accuracy is 89.23% after selecting multiple wavelet bases to extract signal features through wavelet packet analysis and combining with support vector machine.Then,the fusion of several models trained with different wavelet bases to extract parameters improves the accuracy to 91.67%.Octave analysis method is studied,then the signal of the sub-band energy features as audio parameters,according to the fan blades acoustic signal feature selection characteristic parameters,into the support vector machine training model,the resulting test set accuracy is 98.07%,based on the acoustic characteristics of the method was verified defects can effectively identify the fan blade acoustic signal. |
