Arc reflection method（ARM） is a kind of cable fault locating method widely used nowadays. Pulse echoes are subject to interference coming from the arc voltage fluctuation on fault spot when fault cable is tested using arc reflection method. For this reason， it is difficult to locate the fault spot by identifying the disordered features of pulse echoes. This paper analyzes the causes of arc voltage fluctuation when arc is burning as well as the features of the disturbed pulse signals， and then analyzes time-frequency distribution of the disturbed pulse signals using empirical mode decomposition（EMD）. It proves that pulse signals and arc voltage fluctuation received in the same time can be completely separated according to their arithmetical mean characteristics in a specific period of time. A new extraction algorithm of pulse signals is proposed. It can be used for identifying arc reflection cable fault locations based on arithmetical mean characteristics of the distributed pulse signals over a specific period time. An empirical mode decomposition filter is also proposed. This method overcomes the difficulties in choosing signals basic functions and signals decomposition levels. It is fully data-driven， so it has strong adaptive capacity. Results of pulse signals extraction by recorded real test data proved the proposed algorithm.