A novel approach to damage classification based on bispectral analysis and neural network
Abstract »Damage detection and assessment by using Higher Order Statistic Analysis has proved several times to be doable and particularly reliable; indeed, Fourier transform of displacements’ third-order cumulants, also known as Bispectrum, has the advantage of being able to detect non-linearity in the dynamic response of the structural element, while being insensitive to ambient vibrations and Gaussian noise. Thus, asymmetry in the statistic distribution may be easily spotted and related to the damaged conditions, as the majority of common faults, e.g. fatigue cracks, shows bilinear effects.
In this study, firstly a novel approach to damage localisation, resorting to Neural Networks fed with bispectral data, is presented. Afterwards, NNs’ parameters and architecture, as well as several different selections of input data, are investigated in order to maximise its forecast abilities. To validate the introduced approach, a simple finite element model of a 4-meters-long cantilever beam has been built and data have been generated via FE nonlinear analyses performed on it. This model is intended to be a first concept, as generic as possible, of various beam-like structural elements.