The approach of extracting under frequency oscillation modes based on multi-measurement data, has high precision and reliability. However, the huge amount of data and long calculation time have limited the application of the approach. In this paper, a robust online approach based on improved wavelet transform is proposed to extract the parameters of power system dominant oscillation mode from wide-area measurements. The singular value decomposition (SVD) is used to analyze the positive power spectrum matrix of generator electromagnetism power to determine the orders of oscillation modes. To reduce the covariance signals, the SVD is used to diminish the amount of data which is involved in the extraction. Finally, the modal parameters are extracted from each mode of reduced signals using the improved wavelet algorithm in the specified frequency ranges. Through the study of 4-generator 2-area and EPRI-36 test systems, it is verified that the proposed improved wavelet algorithm could extract accurate oscillation parameters with reduced computation data size to increase calculation efficiency.