Nonlinear system identification by means of SVMs: choice of excitation signals

In this work we discuss the application of Support Vector Machines to the problem of identifying a specific class of nonlinear systems, namely Wiener-Hammerstein systems. Only based on a set of Input/Output measurements, a model is built that approximates well the behavior of the considered system. However, care should be taken when designing suitable excitations, as the performance of the proposed approach turns out to be quite sensitive to the nature of the input signal. This sensitivity is studied here by using several datasets, characterised by different excitation signals, in terms of root mean square value, frequency band, spectrum shape and amplitude distribution.