Abstract: In order to rapidly and accurately detect loose parts in the nuclear power plant, an alarming method for detecting the loose parts based on adaptive auto-regressive （AR） model was presented. Use of adaptive AR model tracks the change of background noise and then whitens the signal, thereby enhancing the SNR (signal to noise ratio), and then calculating the RMS of the whitened signal and according dynamic threshold to alarm. Tests were taken by the use of impact signal and noise of Qinshan Nuclear Power Plant. The test results show that the impact signal can be fast detected by using the method when the SNR is low and the noise changes over time.