In order to overcome a problem of a high false alarm rate of a detection result easily caused by the selection of cluster members in the prior art, the invention provides the mobile user abnormal behavior detection method, a sliding window is used to dynamically obtain data to improve the accuracy of user behavior acquisition, a Duun_index concept is introduced after the initial clustering and incremental stage of a traditional FC algorithm, cluster members generated after increment are selected, then selected high quality members are fused by using a voting algorithm to obtain a final result, a correlation matrix is introduced in the similarity comparison with a normal behavior of a user, the change of an average difference degree is used to determine whether the user behavior is normal or not, and thus the purpose of efficient and accurate abnormality detection is achieved.