A fundamental problem in structural damage detection is to define an efficient feature to calculate a damage index. Furthermore, due to perturbations from various sources, we also need to define a rigorous threshold whose overtaking indicates the presence of damages. In this article, we develop a robust damage detection methodology based on principal component analysis. We first present an original damage index based on projection of the separation matrix, and then, we drive a novel adaptive threshold that does not rely on statistical assumptions. This threshold is analytic, and it is based on matrix perturbation theory. The efficiency of the method is illustrated using simulations of a composite smart structure and experimental results performed on a conformal load-bearing antenna structure laboratory test.