The aim of structural health monitoring of civil structure is the early detection of damage to prevent structure failure. But modelling the behaviour of such structure is a very challenging task due to it uniqueness and to the effect of environmental parameters on the dynamic. In this context, the novelty detection approach appears to be well adapted since it avoids the need of prior hypothesis on the nature of the dynamical behaviour, and integrates all variability factors. The work of this thesis has two principal aims. The first one is to quantify the ability of novelty detection to discriminate damage under strong environmental variations and impulse excitation. The second one is to introduce a new damage sensitive feature, referred as Jacobian Feature Vector (JFV). The JFV calculation is based on the reconstructed state space which exploits the progress achieved in the theory of non-linear dynamical systems, also known as chaos theory. The comparison between AR parameters, widely used for time series analysis, and the JFV is carried out on several case studies. One of them is a three storey wooden laboratory structure subjected to strong environmental variations and controlled excitation.\ \indent Both the JFV and AR parameters are estimated on vibration measurements and normalized by the Mahalanobis distance concept. The experimental results show that damage detection is improved if the excitation contains a part of noise in comparison to a random sequence of pulses. Environmental variations generate an important variability of the feature vectors, making difficult the fitting of any statistical model needed for damage discrimination. Only the extremes damage levels are detected in the worst test configuration. The comparison between the feature vectors reveals a higher dispersion of the JFV components, leading to a lower sensitivity to damage, as well as to the effect of temperature. However, a parametric investigation shows that damage detection can be improved if the the selection process of the state-space reconstruction parameters is optimized. Given the limited performances of AR and JFV in the most difficult configurations, an other damage sensitive feature based on cross-correlation of sensor pairs is proposed. In a complex case study which combines environmental variations and impulse excitation, the performances of this feature vector are promising. A discussion about the more rigorous way to divide the undamaged data to apply novelty detection procedure is also proposed. Finally, since the last step of novelty detection is decision making based on statistical modelling of the normalized damage sensitive features, the robustness of several approaches for the setting of the classification threshold is investigated.
