A deep neural network-assisted metamodel for damage detection of trusses using incomplete time-series acceleration

In this article, a deep neural network (DNN)-driven metamodel is first introduced to damage identification of trusses utilizing acceleration signals incompletely measured from limited sensors. This metamodel can automatically learn its damage-sensitive properties itself to construct a better DNN from previously trained poorer ones. Data utilized to build such a model are generated by finite element method (FEM). In which, inputs are the acceleration behavior corresponding to measurement sensors, and outputs are the damage ratios of truss members. The damage site and extent of low-risk members predicted by the current DNN are eradicated by a suggested damage threshold. This helps to dramatically reduce the number of output units to feed into the next DNNs. By repeating such a procedure, the subsequently upgraded DNNs become more precise after several iterations although they only require a simple network architecture, fewer samples, small epochs and less computational effort. To illustrate the efficiency and robustness of the present metamodel, four 2D and 3D trusses with various damage scenarios are investigated. Obtained results indicate that the suggested approach can reliably diagnose both the location and severity of damaged truss members using acceleration behavior measured by a few sensors in a relatively short period of time, even with high noise levels.