An approach based an multi-scale chirplet sparse signal de-composition is proposed to separate the multi-component polynomial phase signals, and estimate their instantaneous frequencies. In this paper, we have generated a family of multi-scale chirplet functions which provide good local correlations of chirps over shorter time inter-val. At every decomposition stage, we build the so-called family of chirplets and our idea is to use a structured algorithm which exploits information in the family to chain chirplets together adaptively as to form the polynomial phase signal component whose correlation with the current residue signal is largest. Simultaneously, the polynomial instantaneous frequency is estimated by connecting the linear frequen-cy of the chirplet fixations adopted in the current separation. Simula-tion experiment demonstrated that this method can separate the com-ponents of the multi-component polynomial phase signals effectively even in the low signal-to-noise ratio condition, and estimate its in-stantaneous frequency accurately.