Abstract Summary
Monitoring the health and working conditions of critical structures and machines can prevent failures and arrange for effective maintenance schedules. Access to vibration data from states with damages is valuable in order to train models that can perform explicit classification of faults in later states. This usually requires expensive and sometimes not feasible destructive tests in various operating conditions. Recent works have employed numerical models in order to simulate such scenarios. However, simulated data carries usually systematic errors in terms of the originally identified structure on the healthy state. In the present work, a test structure of a simple truss is studied with a Finite Element Model for vibration data generation. Training data sets are generated containing different errors in terms of damping, stiffness, and mode shape accordance in order to emulate real application faults. The training data sets are used after in a Convolutional Neural Network for damage classification. It is investigated how different simulation error sources affect the generalization in terms of correct damage case prediction on a reference data set. The results may aid in strategies for focus on specific FE model updating parameters in order to generate reliable simulated training data.
