The invention discloses an
outlier detection method for time-
series data. The method comprises the following steps of: dividing the time-
series data in a training
data set by the day from the Monday to the Sunday, and then clustering; establishing a data
distribution model, under week
granularity, of the time-
series data by using the maximum cluster in each clustering result; according to the data
distribution model, searching all abnormal values in the training
data set, and respectively acquiring the data
distribution model at each time interval; by searching, judging whether a periodic event which occurs with time
granularity, greater than the week
granularity, as a period exists in the abnormal values which accord with the data distribution model at each time interval; if the periodic event exists, recording the periodic event as a class of
special period mode; judging whether the time-series data in a
test data set accords with a week mode, if so, determining that the time-series data is a non-
outlier, otherwise, judging whether the time-series data accords with the
special period mode; and if the time-series data accords with the
special period mode, determining that the time-series data is the non-
outlier, otherwise, determining that the time-series data is an outlier.