Acoustic Detection Of Wind Turbine Blade Faults With Spatial-temporal Joint Processing And Deep Learning

With the increasing global demand for clean energy,wind power generation and related technologies have developed rapidly,and the installed capacity of wind power generators in the world is increasing worldwide.Wind turbine blades operate in harsh outdoor environments for a long time,and are prone to damage.Blade failures account for about 19.4% of wind turbine failures.Moreover,the size of wind turbine blades continues to increase,which increases the risk of failure and costs.The research on the wind turbine blade fault detection is performed based on combination of the blade radiation noise signal processing and the deep learning.In this paper,the microphone array is used to collect the radiated noise generated by the wind turbine without contact,and the received signal is preprocessed in the space domain and the time domain successively to suppress the complex background noise and interference of the wind farm.Experimental results show that spatial-temporal joint processing can effectively suppress background noise and enhance weak signals radiated by faulty blades.On the basis of time-frequency analysis and feature extraction,the graph signal processing and graph feature extraction methods are mainly studied.Short Time Fourier Transform(STFT)is used to extract the non-stationary features of radiation noise,and Mel Filter banks(Fbank)are used to simulate the human ear’s perception of radiation noise.According to the Graph Fourier Transform theory,the signal feature analysis domain is extended from the frequency domain to the graph domain,features of short-time graph Fourier transform(STGFT)and graph Mel filter banks(Gbank)are designed.The feasibility and effectiveness of each feature are verified by the recorded wind turbine blade radiation noise signal combined with the classic Res Net neural network.Combined with the feature extraction methods,a deep learning based acoustic detection framework for wind turbine blade faults is proposed,which integrates frequency and graph domain features,and further introduces the most advanced neural network ECAPA-TDNN and MFA-Conformer under the CNN and Transformer architecture in the field of speaker recognition.In this paper,the performance of each algorithm is compared on the data set constructed by the recorded wind turbine blade radiation noise signal,it has been verified that the best performance can be obtained by combination of Fbank+Gbank fusion feature and MFA-Conformer neural network,which can reach 95.5% accuracy on the test set.Finally,an acoustic wind turbine blade fault detection system is designed and implemented,which integrates spatial-temporal preprocessing module,frequency and graph domain joint feature extraction module,deep learning recognition module,and adapts to various hardware and software.The effectiveness of the system is verified by using the recorded wind turbine blade radiation noise.