Deep neural network applied in calibration of transmission frequency gas-microphone photoacoustic

Abstract

Calibration in gas-microphone photoacoustics is not easily achieved, mainly due to the impossibility of finding two identical microphones needed for the differential set-up (each of them introducing non-linear influence on the recorded PA response). In this work, the methodology is developed which determines the influence of the used microphone, manifested through five characteristic frequencies which are functionally bonded to electronic and geometrical properties of the device. This is accomplished with the implementation of two-layer deep neural networks (fig. 1), enabling the filtration of the measured signal and thus removing the influence of the measurement chain on the photoacoustic response. Case study is done on PA measurements of laser-sintered polyamide (PA12), calibrated onto PA response of aluminum. Analysis of the obtained regression model for the prediction of the microphone parameters of PA response of aluminum is given in table 1. It is proven that this methodology successfully calibrates the measurement of examined samples onto a reference sample response, also filtrated from the measurement chain influence. It is also demonstrated that this procedure expands the frequency range for inverse solving of the PA problem, aiming at the estimation of thermal and optical sample properties, as well as improving their accuracy.