Abstract
This research proposes the feasible methods which aim at mass and location estimation of loose parts in a nuclear power plant. AR(auto-regressive)-based signal whitening filter is used as the pre-filter to improve SNR. Neural network is used to perform mass estimation. It can reduce the uncertainty caused by the violation of basic assumptions and the difficulties of calculating the impact contact time, when it is applied to the real situation. The centroid of frequency spectrum, frequency ratio (FR), and the linear predictive coding (LPC) coefficients are used as features of impact signal, which are the inputs of neural network. In addition, the effect due to the energy and location are analyzed. Based on the experimental data of the nuclear power plant, the average relative error is below 30 %, and the average absolute error is below 130g for both training and test results. As for location estimation, the arrival time difference between two sensors which was estimated by short-time RMS detector is reliable. If only two sensors are used, a roughly impact region can still be predicted by the estimated arrival time difference combined with energy ratio of signals.