Structural Health Monitoring (SHM) can be de ned as the process of acquiring andanalyzing data from on-board sensors to evaluate the health of a structure. Classically, an SHMprocess can be performed in four steps: detection, localization, classi cation and quanti cation.This paper addresses damage quanti cation issue as a classi cation problem whereby each classcorresponds to a certain damage extent. Starting from the assumption that damage causesa structure to exhibit nonlinear response, we investigate whether the use of nonlinear modelbased features increases classi cation performance. A support Vector Machine (SVM) is usedto perform multi-class classi cation task. Two types of features are used as inputs to the SVMalgorithm: Signal Based Features (SBF) and Nonlinear Model Based Features (NMBF). SBF arerooted in a direct use of response signals and do not consider any underlying model of the teststructure. NMBF are computed based on parallel Hammerstein models which are identi ed withan Exponential Sine Sweep (ESS) signal. A study of the sensitivity of classi cation performanceto the noise contained in output signals is also conducted. Dimension reduction of featuresvector using Principal Component Analysis (PCA) is carried out in order to nd out if it allowsrobustifying the quanti cation process suggested in this work. Simulation results on a cantileverbeam with a bilinear torsion spring sti ness are considered for demonstration. Results showthat by introducing NMBF, classi cation performance is improved. Furthermore, PCA allowsfor higher recognition rates while reducing features vector dimension. However, classi ers trainedon NMBF or on principal components appear to be more sensitive to output noise than thosetrained on SBF.