Title:Source Localization Based on Deep Learning of Phononic Modes

Abstract:In this work, an artificial neural network (ANN) is developed for source localization in a 2D stress wave scattering setup. The sensor domain consists of a periodic micro-structured medium, of which the band structure and the eigen-modes are exploited and utilized for detecting the angle of arrival. Modal symmetry breaking is identified at the exceptional point (EP) and critical angles (CAs) in the eigen-wavevector band structure. The eigen-modes switch their energy characteristics and symmetry patterns at these branch points, thus demonstrating strong angle dependence. The eigen-modes may also serve as the basis functions of the scattering wave. Therefore, the scattering response inherently possesses the angle-dependent modal properties as well as the linear subspace generated by them, making it naturally suitable for sensing application. An ANN is trained with randomly weighted eigen-modes to achieve deep learning of the modal features and angle dependence. The training data is derived only based on the modal features of the unit cells. Nevertheless, the trained ANN can accurately identify the incident angle of an unknown scattering signal, with minimal side lobe levels and suppressed main lobe width. The performance of the ANN shows superior performance in comparison with standard delay-and-sum technique of estimating angle of arrival. The proposed application of ANN and micro-structured medium highlights the physical importance of band structure topology and modal shape properties, adds extra strength to the existing localization methods, and can be easily enhanced with the fast-growing data-driven techniques.