For aircraft engineers, detecting abnormalities in a large dataset of recorded flights and understanding the reasons for these are crucial development and monitoring issues. The main difficulty comes from the fact that flights have unequal lengths, and data is usually high dimensional, with a variety of recorded signals. This question is addressed here by introducing a new methodology, combining time series partitioning, relational clustering and the stochasticity of the online self-organizing maps (SOM) algorithm. Our method allows to compress long and high-frequency bivariate time series corresponding to real flights into a sequence of categorical labels, which are next clustered using relational SOM. Eventually, by training SOM with a large number of initial configurations and by taking advantage of the stability of the clusters, we are able to isolate the most atypical flights, and, thanks to discussions with experts, understand what makes a flight an “abnormal” data.