In order to eliminate the outliers in the realtime observation data thus to enhance the precision of the following process, this article, by giving a thorough analysis to the characters of both the realtime observation data and the outliers within them, proposes a selfadaptive fivepoint linear prediction data detection method. Tests based on simulated and observed data show, comparing with prediction and fixed threshold methods, that the proposed method could increase detecting rate of the outliers by 30% while remarkably reduce the false alarm rate, which verifies that the proposed method prevails the traditional methods in the matter of whether reliability or effectiveness, bears fairly strong stability and can effectively eliminate both isolate and continual outliers.