Sensor networks are increasingly applied for monitoring diverse environments and applications. Due to their unsupervised nature of operation and inexpensive hardware used, sensor nodes may furnish readings of rather poor quality. We thus need to devise techniques that can withstand "dirty" data during query processing. In this paper we introduce a robust aggregation framework that can detect and isolate spurious measurements from computed aggregate values. Such readings are not injected in the reported aggregate, in order not to obscure the outcome, but are still maintained and returned to the user/application, which may investigate them further and take appropriate decisions. In addition, our framework provides a form of positive feedback to the user by enhancing the result with a set of nodes that contain the most characteristic values out of those included in the aggregation process. We perform an extensive experimental evaluation of our framework using real traces of sensory data and demonstrate its utility to the monitoring of applications.