A novel multilevel fault alarming filter model is presented by studying the distributions and characteristics of the fault alarms in the radio, switching and transferring of mobile communication networks. Firstly, the fault alarms, which are collected in a realtime way, are filtered and preprocessed to be normalized. Secondly, the fault alarms are coded in binary, and the result data are used by the Gaussian perception network to preclassify them to reduce the alarms that the BP network are supposed to process. Finally, the points of failure are located by the adaptive ability and pattern recognition capability of the BP network. The total system can ensure the network to be stable, efficient, and reliable.