Demodulation of extrinsic Fabry-Pe´rot interferometric sensors for vibration testing using neural networks

Abstract

Strain level measurement on a periodically actuated and instrumented structure can provide information about the health of that structure. A simple demodulation system employing artificial neural networks (ANNs) is analyzed for an extrinsic Fabry-Perot interferometric (EFPI) strain sensor. The harmonic content of the nonlinear sensor output for the sinusoidal strain case is used to predict the maximum strain level. Implementations of the demodulation system are demonstrated for both simulated and experimental data using back-propagation neural networks. The simulation uses the theoretical response of the EFPI sensor and the experiment uses an EFPI-instrumented smart composite beam to obtain training and testing data. Excitation is provided by a piezoelectric actuator operating from 50 Hz to 1 kHz. System performance is compared for two-stage and single-stage networks and for differing types of data preprocessing. The ANN systems successfully extract the signal harmonics and predict the peak strain levels. Data preprocessing using principal component analysis produces the best accuracy. The architecture of an EFPI-based health monitoring system is proposed.