Storage spring is one of the key components of on-load tap changer（OLTC）of power transformer during its operation of voltage regulation

which is closely related to the success of whole switch-over process. According to the chaotic features of vibration signals resulted from the switch-over process

the phase space of vibrations signals is built first. The K-means ++ clustering algorithm is applied to analyze the distribution of phase points in the high dimensional space

where the hierarchical agglomerative clustering is selected to determine the number of clustering.The index of inter class divergence is defined to quantitatively describe the space point distribution of OLTC vibration signals. The vibration signals of the combined OLTC is measured in normal condition and storage spring fault with different abrasion degree. The calculated results have shown that the reasonable number of clustering is important to improve the accuracy of clustering analysis results of vibration signals of OLTC. The inter class divergence can recognize the different abrasion degree of OLTC storage spring

which could provide the important reference for the vibration monitoring technology of OLTC mechanical condition.