Improvements on the offline performance evaluation of fault prognostics methods

The definition of adequate figures of merit for the specification and comparison of failure prognostics methods has presented itself as a challenge for the scientific community. Various metrics have been proposed in the literature for this purpose and the need for standardization becomes increasingly evident. This paper describes improvements on the use of performance evaluation methods based on the probability integral transform (PIT), previously proposed by the authors. The novel contributions described in the paper relate to the definition of critical values and confidence bounds which can provide valuable information for the users of such evaluation methods. One important aspect which is presented is the incorporation of the information of the number of available ground truth data sets into the evaluation. To the best of authors´ knowledge, this is the first time that this relevant aspect is incorporated into prognostics performance metrics.