In the structural health monitoring of bridges, the effects of the operational and environmental variability challenge the reliability of damage detection performance during long-term monitoring. In order to filter those effects to enhance novelty detection on bridges, in the last decade recourse has been made to so-called machine learning algorithms. Generally, such algorithms are based on purely data-based methods because they only rely on features from the structural response data and they are independent of the structural complexity. However, due to their formulation, some of those algorithms most commonly used are only able to remove linear patterns from the features, which might not be appropriate in cases where the structural stiffness varies, for instance, with the temperature gradient range. Therefore, this paper presents an algorithm based on Gaussian mixture models for damage detection under unknown sources of variability and compares it with well-known linear algorithms and one nonlinear algorithm with the capability to model nonlinearities of the baseline data. The study is performed on standard data sets from the Z-24 Bridge, in Switzerland.