In autonomous systems, planning and decision making rely on the estimation of the system state across time, i.e. state tracking. In this work, a preference model is used to provide non ambiguous estimates at each time point. However, this strategy can lead to dead-ends. Our goal is to anticipate dead-ends at design time and to blame root cause preferences, so that these preferences can be revised. To do so, we present the preference-based state estimation approach and we apply a consistency-based meta-diagnosis strategy based on preference relaxation. We evaluate our approach on a robotic functional architecture benchmark.