The need for an active approach to fault tolerance in swarm robotics systems is well established. This will necessarily include an approach to fault diagnosis if robot swarms are to retain long-term autonomy. This paper proposes a novel method for fault diagnosis, based around behavioural feature vectors, that incorporates real-time learning and memory. Initial results are encouraging, and show that an unsupervised learning approach is able to diagnose common electro-mechanical fault types, and arrive at an appropriate recovery option in the majority of the cases tested.