This dissertation used the noise data in vertical recording during 2012 as the research object． We divided the noise records to 5min short segment，calculated the power spectral of each segment，and drew PDF diagram by using probability distribution functions． We then determined the station of the high or low noise reference line using the method of the grid probability． Then on the basis of the anomalies of the station in the PDF，the abnormal noise were divided into four categories: dropped packets abnormal，low noise abnormal，high noise abnormal，median noise abnormal． Four selection methods were found by the station of the high or low noise reference line，and real-time monitoring system of seismic noise was formed by combining four selection methods． Noise records of 85 seismic stations in Fujian Province in July 2013 was selected to verify． The results show that noise-recognition of most station anomalies can reach 99% ． As the average error rate is less than 1%，the effect of selection is very good． Therefore，the method can be applied to station noise in real-time monitoring．