Acoustical Blade Fault Diagnosis Based On Transfer Learning Under Variable Conditions

As the key component of wind turbine,the blade often bears high complex alternating load stress during operation,which is prone to surface peeling,erosion cracks and deformation,resulting in significant economic losses and even casualties.Therefore,it is significant to carry out fault detection for wind turbine blade.Compared with blade fault detection methods such as vibration analysis,infrared imaging and fiber Bragg grating,the blade acoustic detection method based on microphone has the advantages of non-contact measurement,flexible sensor installation and convenient operation and maintenance.However,due to the influence of conditions such as operating environment,blade diameter in practical.The operating acoustic signals of blades collected by different wind farms are quite different,resulting in the performance deterioration of the existing model in the cross-condition diagnosis.Therefore,this paper focuses on the acoustic diagnosis method of blade fault under variable working conditions by introducing the transfer learning technology.The specific contents and results are as follows:(1)By analyzing the spectrum variation characteristics of acoustic signal generation and propagation under variable working conditions,a feature extraction method of voiceprint based on spectrum gravity center and wavelet packet decomposition is proposed.Time-frequency domain analysis based on Short-Time Fourier Transform explores the acoustic signal spectrum characteristics under different working conditions such as blade diameter,drain hole and wind speed.According to the spectrum distribution characteristics of the acoustic signal,the quasi-periodic index based on the spectrum center of gravity is designed to calibrate the endpoints of the effective acoustic pulse fragments in the collected continuous signal.On this basis,the wavelet packet decomposition technology is used to extract the characteristics of the voiceprint energy spectrum that characterize running state of the blade.The measured data verifies the effectiveness of the proposed method.(2)Aiming at the problem that the existing transfer metric methods have insufficient ability to describe the difference of blade sample distribution between the source domain and the target domain in the diagnosis of blade cross-working conditions,a feature transfer metric method integrating the sample density index and the weighted maximum mean discrepancy is proposed.Design a density evaluation index based on the aggregation degree of each category of samples,and combine the label distribution characteristics of the local neighbor circles of the samples,that assign a single sample weight to quantify its contribution to the overall distribution.The mapping feature matrix based on weighted maximum mean discrepancy is iteratively constructed.Finally,the support vector machine technology is used to classify and identify the mapped blade samples under cross working conditions.The simulation results show that the WMMD metric method can more accurately measure the distribution difference between the source domain and the target domain,while has higher recognition accuracy and F1 score than MMD metric.(3)Aiming at the problem of insufficient information of blade running state extracted by traditional signal decomposition technology,a deep transfer learning fault diagnosis method based on local weighted maximum mean discrepancy metric is proposed.Since traditional deep transfer learning network,a weighted maximum mean discrepancy strategy based on local alignment of category subdomains is designed to optimize the existing feature domain self-adaptation method,which only considers the distribution difference between the whole domain,so as to improve the cross-domain recognition effect of blade.In the Softmax classification layer,the label probability distribution of the source domain and the target domain is minimized,and the pseudo-label information of the target domain is optimized.The experiments show that the optimized deep transfer learning network model has higher recognition accuracy in the fault diagnosis of wind turbine blades across working conditions.