The loose parts monitoring system is one of the important instrument and control systems in nuclear power plants, which is used to detect metallic loose parts within the reactor coolant system. However, this system now and then is interfered by the strong background noises on account of flow-induced vibrations and equipment operations. Therefore, high fault alarm rate is inevitable and ubiquitous. The proposed technique uses big data and artificial intelligence to analyze impact signals in time-frequency domain. In addition, a method of spectral feature extraction is developed and programmed as a real-time fault diagnosis platform. Through long-term practical implementation, the result shows it can effectively discriminate the real metal impact alarms from false ones more than 90% to ensure operation security of the nuclear power plant. Furthermore, this technique can be applied to the fault prognosis in other kinds of power plants.