Résumé : This paper addresses the topic of condition monitoring of wind turbine blades and presents a learning-based approach to fault detection. The proposed scheme utilises Principal Components Analysis and Autoencoders to derive data-driven models from root-bending moment and other measurements. The models are trained with real data obtained from a fault-free wind turbine, and then validated on data corresponding to unknown health condition. Online test statistics, employing static thresholds and Generalized Likelihood Ratio tests, are used on residual signals generated by discrepancies between the actual and reconstructed measurements to detect deviations from nominal operation. The efficacy and effectiveness of the proposed methods are demonstrated using real-life data collected from wind turbines experiencing blade faults.