Abstract: In the determination of the gamma air kerma conventional true value (CAK) in a minitype reference radiation (MRR), the scattering γ spectrum is affected by the radioactive statistical fluctuation and other electronic noise. Moreover, the excess data of the spectrum after dispersing also make it unable to characterize the MRR accurately and decrease the construction efficiency and prediction precision of the prediction model of CAK. The wavelet analysis and principal component analysis (PCA) were employed to de-noise the scattering γ spectrum and extract the metrological feature components. When the metrological feature components were used to replace the spectrum for CAK prediction model construction, the construction efficiency and prediction precision were improved significantly.