The invention discloses a time sequence invariant feature extraction method for a high-order dynamic function connection network, and the method mainly comprises the steps: (1) segmenting a whole magnetic resonance sequence into a plurality of sub-sequences for a given sliding window width and step length; (2) calculating the correlation between the brain regions in each sub-sequence to obtain a dynamic function connection network, and further calculating the central moment characteristics of the dynamic function connection sequences of any two brain regions; and (3) regarding the connection sequence of each brain region and other brain regions as a one-dimensional random sequence for each sub-sequence of the dynamic function connection network, calculating the correlation of any pair of brain regions, and further constructing a high-order dynamic function connection network and obtaining the central moment characteristics of the high-order dynamic function connection network. According to the method, the central moment characteristics are used as the time sequence invariant characteristics of the dynamic function connection network and the high-order dynamic function connection network, the deep association relationship of brain function connection can be captured, and stable characteristics are provided for subsequent medical image processing.