An extended classifiability index for feature selection in nuclear transients

Abstract A preliminary step in transient classification for diagnostic purposes is the identification of those measured plant parameters (features), which are most sensitive to the faults and malfunctions and thus can be used most effectively for their classification. This is particularly important for nuclear power plants, where hundreds of parameters are monitored for operation and safety reasons so that expert judgment alone cannot effectively drive the feature selection. Moreover, the sensitivity of the particular transient classification technique to the different plant features must be considered for higher classification efficiency. In this paper, a feature selection algorithm is proposed based on the extension to the transient case of a classifiability index which can be directly computed from the plant measured data and used to filter out irrelevant or redundant features.