MICROPHONE ARRAY TECHNIQUES FOR TRAILING EDGE NOISE BASED ON CONVOLUTIONAL NEUTRAL NETWORK

Trailing edge noise is a major contribution to the airframe noise and has been studied in various processing methodologies such as wavelet analysis and velocity field comparison. To further understand the noise mechanism and improve the result quality in wind tunnel, microphone array technique has been applied to obtain the sound source distribution by different traditional processing algorithms such as beamforming algorithms and deconvolution algorithms. Rather than solving the sound transfer function, a convolutional neutral network (CNN) method has been proposed based on machine learning for complex objects, and proven efficient in limited number of incoherent monopole sources. However, trailing edge noise is an emission of distributed sound sources and most phased array methods including the CNN method assume the monopole sources. In this study, CNN method is investigated from monopole sources to simple line sources in order to find out an applicable condition for trailing edge noise. By using different definitions of accuracy and loss, the network is trained, tested, and compared with other researchers' results to validate on monopole sources. Then a new test strategy has been proposed and proven helpful in understanding the wrong prediction of the CNN model. The confidence level of the applied region and the method to use the prediction results are proposed to standardize and normalize the applicable range of the phased array. Then, a new training strategy is proposed by adding statistical parameters of the line sources as training data to get a direct source strength distribution of line sources. An experiment of the trailing edge noise of NACA0012 airfoil is conducted in Beihang D5 aeroacoutics wind tunnel to validate the methods. These results indicate that CNN method might be able to detect a more general feature of the sound source distributions, which might be helpful in constructing a network with a statistical output.