Acoustic source localization by deep-learning attention-based modulation of microphone array data

We proposed a deep-learning attention-based methodology to predict acoustic sources obtained from pendulum impact experiments using the Cluster-Self Adaptive Network (CSAN) and showed that the experimental data required for training can be reduced by 50% without losing significant localization accuracy. Acoustic signals due to pendulum impacts on a homogeneous steel plate were recorded by an asymmetric microphone array. Important wavelet features were extracted by transforming the acoustic signals using continuous wavelet functions and reduced the data dimensionality by principal component analysis. Two data sampling strategies (random and Latin hypercube) were investigated to study the effect of the density of training domains on the model performance. The attention-based modulation strategy was employed on microphone positions for data augmentation and prediction of acoustic sources. A comprehensive analysis of the CSAN-based localization results including error estimation was performed. The outcome was contrasted against delay-and-sum beamforming localization results. • A localization method for acoustic sources using artificial neural networks was presented • An attention-based Cluster-Self Adaptive Network (CSAN) was applied for modulating the different weighting factors among the recorded signals • Feature engineering techniques were applied to optimize the model workflow • Error estimation was performed to demonstrate the generalizability of the CSAN in localizing acoustic sources • Human effort to perform physical experiments could be reduced by 50%