| Due to the invasion of dust erosion and lightning strikes in natural environment,the surface of wind turbine blades is prone to damage,and online fault detection is crucial in particular.Compared with contact blade fault detection methods such as vibration,acoustic emission and infrared,the non-contact acoustic detection method has the advantages of flexible sensor installation and convenient maintenance.Nevertheless,the traditional voiceprint characterization method of non-contact detection encounter disadvantages of readily coupling disturbance noise,static voiceprint limitations,and the diagnostic model is affected by sample imbalance.From the perspective of ear hearing separable,this paper proposes to explore a voiceprint characterization and recognition method of the wind turbine blade based on the spectrogram.The specific contents and results are as follows:(1)To solve the problem that the original operating sound signal of blade is susceptible to noise coupling interference,a static voiceprint characterization method(Auditory Perception Wavelet Packet Coefficient,APWPC)based on auditory perception wavelet packet transform is established.By simulating the masking characteristics of human ear hearing,the auditory perception wavelet packet decomposition tree construction algorithm is designed to perform degraded pruning processing,which realizes the shielding of interference components and capture of auditory separable effective information.Via analyzing the random distribution characteristics of noise interference,a non-overlapping perceptual sub-band voiceprint feature characterization method with threshold adaptive noise reduction is designed to improve the anti-noise ability of auditory perceptual features.The simulation results show that APWPC performs best compared with Mel cepstrum coefficient feature and wavelet packet energy feature,and the recognition accuracy of classification model reaches 93.8%.(2)To break the limitation of single static voiceprint feature description,a blade dynamic voiceprint feature fusion algorithm(Fisher Dynamic Fusion Voiceprint,F-DFV)with homologous variable scale is proposed.Combined with the similarity of impeller rotation and periodic steady-state characteristics,the continuous frame difference feature in time-space and the enhanced spectrogram homomorphic similarity algorithm are designed to characterize the dynamic voiceprint characteristics of the operating blade.Fisher discriminant criterion was introduced to integrate homologous multi-scale dynamic and static voiceprint features,which can comprehensively characterize blade health status.The simulation results show that recognition accuracy of the model based on F-DFV is increased by 5.7%compared with APWPC.(3)Aiming at the problem that the blade acoustic diagnosis model is seriously affected by sample imbalance,a blade status recognition method based on global cost-sensitive dictionary learning(G-CSDL)is proposed.A minority class sample synthesis method based on regional triangle midline is designed to accomplish pre-equalization of classes in feature space,which can reduce the influence of model skew.In accordance with the sparse coding importance of scarce,heterogeneous and overall,the global cost-sensitive coefficient of class equilibrium is constructed to enhance the agility of the model to fault classes.The simulation results show that fault missed diagnosis rate of G-CSDL is reduced to 1.3%. |
