A Methodology for Diagnosis of Damage by Machine Learning Algorithm on Experimental Data

Abstract

For optimal performance, the machine learning algorithms requires enough number of training data. The major issue with these algorithms in damage diagnosis is unavailability of data contains range distinct damaged scenario recorded from a real structure. The generation of such data experimentally is a challenging task. To address this issue, a methodology is proposed by which the diagnosis algorithm is trained with numerically simulated database and tested on real experimental data. The proposed methodology consists a unique feature extraction procedure that extract the features from simulated data which are close to experimental data. The methodology is employed for quantification and localization of debonding in a metallic stiffened panel using the vibration-based approach. The natural frequency of the undamaged numerical simulated model is experimentally validated; the validated model used to create numerically simulated damage database. The feature is extracted from first mode displacement raw data. The Artificial Neural Networks (ANNs) have been chosen as diagnosis algorithm. The algorithm structure has been optimized on numerically simulated database. Finally, the feasibility of the algorithm is tested by processing real experimental data trough optimized algorithm recorded by Laser Doppler Vibrometer (LDV). The proposed methodology is produced a good prediction accuracy in quantification and localization of debonding on experimental data.