In order to realize intelligent diagnosis of high voltage shunt reactor, a fault diagnosis method based on cross-wavelet transform is proposed. The vibration signal test platform is built to collect the reactor vibration signals under different conditions. The cross-wavelet power spectrum of reactor under three states is compared and analyzed. The reactor faults are qualitatively analyzed by the changes of color, saliency level curve and phase angle in the spectrum. The characteristic frequency band of the signal is determined and the characteristics are extracted. In frequency band, RGB parameters and phase angle information are used to construct feature matrices. Finally, matrix recognition is used to measure the difference between different feature matrices. The results show that the cross-wavelet power spectrum has excellent noise stability and can accurately reflect the correlation of reactor vibration signals in different frequency bands. With the decrease of compression force, the correlation among D2, D3 and D4 bands decreases, while the correlation between D6 bands increases. In addition, the characteristic matrix of reactor under different states has obvious difference, and the matrix similarity index can quantify the difference in the cross-wavelet spectrum after loosening.