Experimental validation and performance analysis of deep learning acoustic source imaging methods

Deep Neural Network (DNN) models offer a very attractive alternative to existing acoustic source imaging techniques, such as acoustic beamforming, due to their ever-growing potential with increasing computational power. Source resolution of acoustic beamforming methods can be limited at relatively low frequencies and despite the use of deconvolution methods, the source maps may also possess sidelobes, particularly at high frequencies, and main lobe smearing. Since the application of DNN models to acoustic source imaging problems is a very recent concept, there are little data available regarding the robustness and performance analysis of DNN models. In this paper, a numerical DNN model for acoustic source imaging is presented, that is trained using random phase pressure data generated from six sources over a series of design frequencies, ranging from 1000 Hz to 20,000 Hz. The DNN model robustness is tested, by including extraneous Gaussian white noise and tonal noise inputs near the design frequency. The DNN models are also tested at frequencies that slightly differ from the design frequencies, thus calculating a frequency range over which the DNN model can generate adequate acoustic source estimation. The DNN models are also tested using different number of sources that what they are trained for, to further test robustness. An experimental validation is conducted using a single speaker that is systematically placed over a speaker grid to generate training data via acoustic superposition. The performance of the experimentally trained DNN model, albeit in its infancy, shows exceptional noise source localization capability and a very promising start for a more sophisticated experimentally trained DNN model suitable for aeroacoustic testing in a wind tunnel facility.