Abstract: To address the insufficient stability of modal identification caused by environmental background noise, low-frequency disturbances, and unit-operation noise in vibration monitoring signals of high arch dams, this study conducts a denoising analysis of vibration data from XLD arch dam for modal identification. Considering the lack of clean reference signals in field measurements, finite element simulated seismic responses are used as baseline signals. White noise and pink noise are superimposed to construct contaminated signals at four noise levels of 20%, 40%, 60%, and 80%. The denoising performance of variational mode decomposition (VMD), singular value decomposition (SVD), discrete wavelet transform (DWT), and ensemble empirical mode decomposition (EEMD) is then compared. The results show that VMD achieves higher signal-to-noise ratios and lower root mean square errors in most cases, indicating better overall performance. SVD performs stably under white noise, whereas DWT and EEMD are more sensitive to noise type and noise level. After applying VMD to measured seismic and ambient vibration data, the main time-frequency characteristics are generally preserved, while non-dominant frequency components are suppressed. Low-frequency noise frequencies misidentified in seismic responses are reduced, and low-frequency scattered poles in ambient vibration stabilization diagrams are decreased, thereby improving the stability of modal identification.