Study on sound source localization inside a structure using a domain transfer model for real-world adaption of a trained model

Abstract

Sound source localization (SSL) is important to reduce noise of products such as machinery and electrical appliances. Currently, SSL methods have been proposed that use the correlation of time-frequency signals of sound waves observed by multiple microphones. However, the application of these SSL methods are limited to the case where the acoustic signals can be directly observed. In the case of estimating the location of a sound source inside a structure from outside the structure, these methods are not applicable because the acoustic signal is observed as indirect sound. An SSL method using deep neural network (DNN) and computer-aided engineering (CAE) was proposed to estimate sound sources inside structures. This method successfully estimates the location of sound sources inside a structure from signals observed by accelerometers installed on the outer surface of the structure in both CAE and real domains. Nevertheless, the proposed method still has the challenge of adapting DNNs trained in CAE domain to real domain. In this research, deep convolution auto-encoder (DCAE) was used as a domain transformation model built from CAE and actual experimental data. DCAE contributed to the improvement of the SSL performance of the trained DNN over the baseline.