On-line diagnosis must accommodate the existing sensoring capabilities of a system, which often results in limited diagnosability. However, although faults may not be always discriminable, there are generally operating modes of the system in which they are. Active diagnosis relies on applying specific inputs to the system so as to exhibit additional symptoms that help refining the diagnosis. The idea of this paper is to use the diagnosability properties to drive the system towards modes with increased diagnosability with respect of safety considerations. A new finite state machine called the active diagnoser is defined by abstracting continuous dynamics and taking into account controllability and safety constraints. The active diagnosis problem is then formulated as a conditional planning problem. Hence, the active diagnoser is transformed in an AND-OR graph and active diagnosis plans are computed by an appropriate graph exploration algorithm.