Abstract: A new approach using the wavelet packet transform is proposed for power spectral analysis of stochastic signals. The background component and the resonant component are separated successfully using the proposed approach. The accuracy of frequency detection and the resolving power of the proposed approach are investigated through a spectral analysis of a double-frequency signal. Then the approach is applied to compute the power spectra of the wind-induced response of a large-span roof structure. And the accuracy of spectral estimation for stochastic signals is verified. The numerical results indicate: 1) the proposed approach is efficient and accurate with high resolving power, so it is applicable for power spectrum estimation of various response signals of large roofs induced by the wind; 2) the wavelet packet decomposition of the fluctuating response signal is an effective approach to obtain time histories of the background response and the resonant response.