A distributed particle filtering approach is proposed for fault detection in dynamic systems, where an interacting multiple model particle filter is used at each sensor for joint discrete mode (denoting normal or faulty situations) and continuous state tracking. Adaptive approximations of local state posterior distributions by histograms are aggregated to obtain the final decision regarding the system mode. By increasing the number of sensors, the uncertainty associated with the system mode reduces. Due to communication constraints, the number of bins used for histogram approximation is dynamically adjusted according to the mode distribution. Thus, better performance can be achieved compared with systems without this adaptation.