Since systems are prone to faults, fault detection and isolation are essential activities to be considered in safety-critical applications. In this direction, the availability of a sound estimate of the time instant the fault occurred is a precious information that a diagnosis system can fruitfully exploit, e.g., to identify information consistent with the faulty state. Unfortunately,any fault-detection system introduces a structural delay that,typically, increases in correspondence of subtle faults (e.g., those characterized by a small magnitude) with a consequence that the fault-occurrence time is overestimated. In this paper we propose an ensemble approach to estimate the time instant a fault occurred. We focus on systems that can be described as ARMA models and faults inducing an abrupt change in the model coefficients.