Abstract: Outlier detection efficiency in data stream can be influenced by concept drift when there is noise．An outlier dynamic detection method based on constructive neural networks incremental learning was presented to solve this problem．The outline of the data in the sliding window is acquired,and the learning model is modified．On the other hand,data moving speed and flux can influence the efficiency as well．In this paper,the method of granular analysis was used to improve our method．The best analysis granularity in suitable sliding window was set to find outlier accurately．The simu-late experiment and the experiment of outlier detection in radio communication demonstrated the efficiency of this method．

Key words: Data stream,Outlier detection,Dynamic detection,Constructive neural networks