Phase-sensitive Optical-Time-Domain Reflectometry (OTDR) system is a typical distributed fiber-optic sensing technology to detect and locate multiple dynamic disturbances from the outside, which provides a cost-effective and highly sensitive solution especially for monitoring long or ultra-long perimeters. However, the system is liable to be interfered by laser frequency drifts and environmental noises due to its phase sensitivity. The fluctuant and time-varying backgrounds severely obscure real intrusion signals, which always cause bad detection results or high Nuisance Alarm Rates (NARs). In this paper, an effective signal separation method is proposed to extract true intrusion information from the complicated noisy backgrounds of phase-sensitive OTDR system. The sensing signal in the time-domain at each spatial point is obtained by accumulating the changing trails at different moments. Multi-scale wavelet decomposition is employed on the temporal signal to get the detailed components at different scales. By selectively recombining the scale components, it can easily extract the real intrusion signal, and separate the fluctuant frequency-drift induced phase noises, and the time-varying sound or other interferences caused by the air movements, which are respectively located at different time-frequency components. Moreover, the experimental results show that the event type could be divided and discerned from the time-frequency energy distribution at different scale. Thus nuisance and false alarms in practical applications of phase-sensitive OTDR system can be decreased significantly by this way of signal separation and extraction. This technique provides a useful solution for the intrusion detection and identification of the phase-sensitive OTDR in complicated environments, and paves the way for many important applications such as long perimeter security, oil or gas pipe safety monitoring, large-scale structure health detection and fault diagnosis and so on.