Neural network data analysis for laser-induced thermal acoustics

Abstract

A general, analytical closed-form solution for laser-induced thermal acoustic (LITA) signals using homodyne or heterodyne detection and using electrostrictive and thermal gratings is derived. A one-hidden-layer feed-forward neural network is trained using back-propagation learning and a steepest descent learning rule to extract the speed of sound and flow velocity from a heterodyne LITA signal. The effect of the network size on the performance is demonstrated. The accuracy is determined with a second set of LITA signals that were not used during the training phase. The accuracy is found to be better than that of a conventional frequency decomposition technique while being computationally as efficient. This data analysis method is robust with respect to noise, numerically stable and fast enough for real-time data analysis.