基于深度神经网络的声源成像方法研究

阵列信号处理方法是麦克风阵列声源定位的核心，波束成形方法被广泛用于声源成像，传统波束成形方法通常在低频时受限于瑞利分辨率极限，在高频时会产生较高的旁瓣幅值。随着硬件计算能力的快速提高，以深度学习为主的人工智能热潮再次涌现，深度学习显著提高了多个领域的极限性能，也给气动声学成像提供了新的解决方法。近年来，基于深度神经网络的声源成像方法具有优于传统方法的成像性能，但同时缺乏对深度神经网络模型泛化能力的系统认识与实验验证。本文分别基于模拟声源数据和实验声源数据对深度神经网络模型的鲁棒性进行系列研究，结果表明，DNN模型具有比较好的鲁棒性，同相位多点声源模型的鲁棒性优于随机相位多点声源模型，高频模型的鲁棒性优于低频模型；基于实验声源数据的模型更容易训练，且低频和高频之间的训练收敛速度以及成像效果差异很小，在与训练实验数据同类型的测试数据上能得到非常完美的成像效果；训练数据与测试数据之间的差异，会对成像效果造成较大的影响，但模型对不同于训练声源的其他声源仍然具有成像潜力，研究结果证明了深度神经网络用于真实多声源成像的有效性和先进性。

[1] MOSHER M, WATTS M, BARNES M, et al. Microphone array phased processing system(MAPPS): phased array system for acoustic measurements in a wind tunnel[M]. SAE International, 1999.
[2] SIJTSMA P. Phased array beamforming applied to wind tunnel and fly-over tests[M]. SAEInternational, 2010.
[3] DOUGHERTY R. Beamforming in acoustic testing[M]. Springer Berlin Heidelberg, 2002:62-97.
[4] CHIARIOTTI P, MARTARELLI M, CASTELLINI P. Acoustic beamforming for noise sourcelocalization–Reviews, methodology and applications[J]. Mechanical Systems and Signal Processing, 2019, 120: 422-448.
[5] SIJTSMA P. CLEAN based on spatial source coherence[J]. International Journal of Aeroacoustics, 2007, 6(4): 357-374.
[6] BROOKS T, HUMPHREYS W. A deconvolution approach for the mapping of acoustic sources(DAMAS) determined from phased microphone arrays[J]. Journal of Sound and Vibration,2006, 294(4-5): 856-879.
[7] HINTON G, SALAKHUTDINOV R. Reducing the dimensionality of data with neural networks[J]. Science, 2006, 313(5786): 504-507.
[8] KRIZHEVSKY A, SUTSKEVER I, HINTON G. ImageNet classification with deep convolutional neural networks[J]. Communications of the ACM, 2017, 60(6): 84-90.
[9] SILVER D, HUANG A, MADDISON C, et al. Mastering the game of Go with deep neuralnetworks and tree search[J]. Nature, 2016, 529(7587): 484-489.
[10] LECUN Y, BENGIO Y, HINTON G. Deep learning[J]. Nature, 2015, 521(7553): 436-444.
[11] BIANCO M J, GERSTOFT P, TRAER J, et al. Machine learning in acoustics: a review[J].arXiv preprint arXiv:1905.04418, 2019.
[12] VERA-DIAZ J, PIZARRO D, MACIAS-GUARASA J. Towards end-to-end acoustic localization using deep learning: from audio signals to source position coordinates[J]. Sensors, 2018,18(10): 3418.
[13] PUJOL H, BAVU E, GARCIA A. BeamLearning: An end-to-end deep learning approach forthe angular localization of sound sources using raw multichannel acoustic pressure data[J]. TheJournal of the Acoustical Society of America, 2021, 149(6): 4248-4263.
[14] KUJAWSKI A, HEROLD G, SARRADJ E. A deep learning method for grid-free localizationand quantification of sound sources[J]. The Journal of the Acoustical Society of America, 2019,146(3): 225-231.
[15] LEE S Y, LEE S, JUNG J H. Acoustic source localization for single point source using convolutional neural network and weighted frequency loss[C]//INTER-NOISE and NOISE-CONCongress and Conference Proceedings: volume 261. 2020: 5674-5681.
[16] HUANG G, LIU Z, VAN DER MAATEN L, et al. Densely connected convolutional networks[C]//2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2017.
[17] LEE S Y, CHANG J, LEE S. Deep learning-based method for multiple sound source localizationwith high resolution and accuracy[J]. Mechanical Systems and Signal Processing, 2021, 161:107959.
[18] LEE S Y, CHANG J, LEE S. Deep learning-enabled high-resolution and fast sound sourcelocalization in spherical microphone array system[J]. IEEE Transactions on Instrumentationand Measurement, 2022, 71: 1-12.
[19] WAGNER G P, BAUERHEIM M, PARISOT-DUPUIS H. Deconvoluting acoustic beamformingmaps with a deep neural network[J]. INTER-NOISE and NOISE-CON Congress and Conference Proceedings, 2021, 263(1): 5397-5408.
[20] CASTELLINI P, GIULIETTI N, FALCIONELLI N, et al. A neural network based microphonearray approach to grid-less noise source localization[J]. Applied Acoustics, 2021, 177: 107947.
[21] MA W, LIU X. Phased microphone array for sound source localization with deep learning[J].Aerospace Systems, 2019, 2(2): 71-81.
[22] XIN C, WANG D, YIN J X, et al. A direct position-determination approach for multiple sourcesbased on neural network computation[J]. Sensors, 2018, 18(6): 1925.
[23] YANGZHOU J Y, MA Z Y, HUANG X. A deep neural network approach to acoustic sourcelocalization in a shallow water tank experiment[J]. The Journal of the Acoustical Society ofAmerica, 2019, 146(6): 4802-4811.
[24] NIU H Q, REEVES E, GERSTOFT P. Source localization in an ocean waveguide using supervised machine learning[J]. The Journal of the Acoustical Society of America, 2017, 142(3):1176-1188.
[25] NIU H Q, GONG Z X, OZANICH E, et al. Deep-learning source localization using multifrequency magnitude-only data[J]. The Journal of the Acoustical Society of America, 2019,146(1): 211-222.
[26] OZANICH E, GERSTOFT P, NIU H Q. A feedforward neural network for direction-of-arrivalestimation[J]. The Journal of the Acoustical Society of America, 2020, 147(3): 2035-2048.
[27] LIU Y N, NIU H Q, LI Z L, et al. Deep-learning source localization using autocorrelation functions from a single hydrophone in deep ocean[J]. JASA Express Letters, 2021, 1(3): 036002.
[28] XU P W, ARCONDOULIS E, LIU Y. Acoustic source imaging using densely connected convolutional networks[J]. Mechanical Systems and Signal Processing, 2021, 151: 107370.
[29] SARRADJ E. Three-dimensional acoustic source mapping with different beamforming steeringvector formulations[J]. Advances in Acoustics and Vibration, 2012, 2012: 1-12.
[30] HEROLD G, SARRADJ E. Performance analysis of microphone array methods[J]. Journal ofSound and Vibration, 2017, 401: 152-168.
[31] DOUGHERTY R P. Functional beamforming[C]//Proceedings on CD of the 5th Berlin Beamforming Conference. 2014.
[32] SARRADJ E, SCHULZE C, ZEIBIG A. Identification of noise source mechanisms using orthogonal beamforming[C]//Noise and Vibration: Emerging Methods. 2005.
[33] MERINO-MARTíNEZ R, SIJTSMA P, SNELLEN M, et al. A review of acoustic imagingmethods using phased microphone arrays[J]. CEAS Aeronautical Journal, 2019, 10(1): 197-230.
[34] BROOKS T, HUMPHREYS W. Extension of DAMAS phased array processing for spatialcoherence determination (DAMAS-C)[M]. 2006: 2654.
[35] DOUGHERTY R P. Extensions of DAMAS and benefits and limitations of deconvolution inbeamforming[M]. 2008: 12.
[36] BRUSNIAK L. DAMAS2 validation for flight test airframe noise measurements[J]. BerlinBeamforming Confer, 2008.
[37] HECHT-NIELSEN R. Theory of the backpropagation neural network[J]. Neural Networks,1988, 1: 445.
[38] RUMELHART D, HINTON G, WILLIAMS R. Learning representations by back-propagatingerrors[J]. Nature, 1986, 323(6088): 533-536.
[39] NESTEROV Y. Gradient methods for minimizing composite functions[J]. Mathematical Programming, 2012, 140(1): 125-161.
[40] PASCANU R, MIKOLOV T, BENGIO Y. On the diffculty of training recurrent neural networks[C]//International conference on machine learning. 2013: 1310-1318.
[41] DUCHI J, HAZAN E, SINGER Y. Adaptive subgradient methods for online learning andstochastic optimization[J]. Journal of Machine Learning Research, 2011, 12(61): 2121-2159.
[42] ZEILER M D. Adadelta: an adaptive learning rate method[J]. arXiv preprint arXiv:1212.5701,2012.
[43] KINGMA D P, BA J. Adam: A method for stochastic optimization[J]. arXiv preprintarXiv:1412.6980, 2014.
[44] SRIVASTAVA N, HINTON G, KRIZHEVSKY A, et al. Dropout: a simple way to preventneural networks from overfitting[J]. The journal of machine learning research, 2014, 15(1):1929-1958.
[45] LECUN Y, BOTTOU L, BENGIO Y, et al. Gradient-based learning applied to document recognition[J]. Proceedings of the IEEE, 1998, 86(11): 2278-2324.
[46] SHEN W B, SETHI G, ZAKKA K, et al. CS231n Convolutional neural networks for visualrecognition[EB/OL]. http://cs231n.github.io/convolutional-networks.
[47] YANG Y N, LIU Y, LIU R, et al. Design, validation, and benchmark tests of the aeroacousticwind tunnel in SUSTech[J]. Applied Acoustics, 2021, 175: 107847.
[48] BROOKS T F, HUMPHREYS W M, PLASSMAN G. DAMAS processing for a phased arraystudy in the NASA langley jet noise laboratory[C]//16th AIAA/CEAS Aeroacoustics Conference. 2010: 37-80.