This paper presents a new mass and energy estimating method for loose parts (LPs) combining the Karhunen-Loève (K-L) transform and neural network theories in the frequency domain. The detection of LPs was performed using simulated acoustic sensors mounted on the wall of a simulator of a reactor vessel. The impact events were simulated by simple pendulums. The data sampled in the time domain was changed to power spectral densities in the frequency domain using the fast Fourier transform. Then, the K-L transform was used to compress the original information. The final feature space's dimensions can be much less than the original ones. And, the original information remains as much as possible. The experiment showed that the impact characteristics of the LPs could be exactly depicted in the compressed feature space. The calculated mass values were approximately equal to the actual ones.