Neural networks for output-only parameter identification

The presented methodology is composed by two main steps, both of which heavilyemploy neural networks, though in different forms.The problem addressed is the parameter identification of a FEM model of a realstructure, of which acceleration records (usually due to environmental loads) areavailable, but scarce information is available about the forcing process.The first step is mainly a signal analysis: acceleration records are processed in thefrequency domain by means of the FFT algorithm in order to obtain the spectral tensorof the recorded signals. Subsequently, the main features of the spectral tensor areexamined, namely the presence of peaks in the auto-spectral functions, the norm of thecoherence tensor, the norm of the phase angle tensor and the possibility of decomposingthe spectral tensor in the tensor product of an unknown vector by itself