Fundamental study on sound source localization inside a structure using a deep neural network and computer-aided engineering

Abstract

This study proposes a sound source localization (SSL) method applicable to sources inside structures such as mechanical equipment or buildings. Presently, an SSL system employing a microphone array based on the time difference of arrival estimation can be used to localize a sound source in the same acoustic space as the microphone. However, conventional SSL methods cannot be adopted when the sound source is located inside a structure. Achieving SSL is more difficult in the case of an indirect sound than in that of a direct sound, because the correlation between the observed signals becomes stronger owing to the effect of coupling between the acoustics and structure. To solve this problem, an SSL method employing a deep neural network and computer-aided engineering, which is applicable to the structure’s interiors, is proposed. The proposed method’s effectiveness and feasibility are examined via numerical and actual experimentation. The proposed method can estimate the position of the sound source inside the structure based on the spectrum measured by an accelerometer on the surface of the structure. The results of the numerical experiment indicate a test accuracy of 93.20%, whereas the actual experiment yielded an accuracy of 61.53%. The learning and validation curves show that the accuracy of the actual experiment is lower owing to the occurrence of overlearning, which results from the small amount of data applied. To overcome this issue, data augmentation was used; consequently, the accuracy was improved to 99.82%.